as_thorbecke_19970121.

as_thorbecke_19970121.
Common Focus Point
Technology
Proefschrift
ter verkrijging van de graad van doctor
aan de Technische Universiteit Delft,
op gezag van de Rector Magnificus,
Prof. dr. ir. J. Blaauwendraad
in het openbaar te verdedigen
ten overstaan van een commissie,
door het College van Dekanen aangewezen,
op dinsdag 21 januari 1997 te 10:30 uur
door
Jan Willem THORBECKE
mijnbouwkundig ingenieur
geboren te Oostzaan
Dit proefschrift is goedgekeurd door de promotor:
Prof. dr. ir. A. J. Berkhout
Samenstelling promotiecommissie:
Rector Magnificus, voorzitter
Prof. dr. ir. A. J. Berkhout, (TU-Delft, Technische Natuurkunde), promotor
Prof. dr. ir. J. T. Fokkema, (TU-Delft, Technische Aardwetenschappen)
Prof. dr. ir. P. M. van den Berg, (TU-Delft, Electrotechniek )
Prof. dr. ir. I. T. Young, (TU-Delft, Technische Natuurkunde)
Prof. dr. J. C. Mondt, (U-Utrecht, Aardwetenschappen)
Dr. ir. C. P. A. Wapenaar, (TU-Delft, Technische Natuurkunde)
Dr. ir. W. E. A. Rietveld, (Amoco, Exploration and Production Technology Group)
ISBN 90-9010-123-3
c
Copyright 1997,
by J. W. Thorbecke, Laboratory of Seismics and Acoustics. Faculty of
Applied Physics, Delft University of Technology, Delft, The Netherlands.
All rights reserved. No part of this publication may be reproduced, stored in a retrieval
system or transmitted in any form or by any means, electronic, mechanical, photocopying,
recording or otherwise, without the prior written permission of the author J. W. Thorbecke,
Faculty of Applied Physics, Delft University of Technology, P.O. Box 5046, 2600 GA Delft,
The Netherlands.
SUPPORT
The research for this thesis has been financially supported by the DELPHI consortium.
Typesetting system: LATEX2e
Cover: ’combined snapshots from one point source’
Printed in The Netherlands by: Beeld en Grafisch Centrum Technische Universiteit Delft.
’Alles komt voort uit de chaos’, wordt er in de Griekse Anthologie gezegd. En inderdaad, alles komt voort uit de chaos. Buiten de wiskunde, die enkel met dode
getallen en loze formules te maken heeft en derhalve volmaakt logisch kan zijn, is de
wetenschap niet meer dan een kinderspel in de schemering, het willen vangen van
vogelschaduwen en het willen tegenhouden van de schaduw van gras dat waait in de
wind. (290 [479], Pessoa (1990))
Preface
In January 1991 I finished my MSc project, which was supervised by professor
Fokkema and professor Van Den Berg, at the department of applied Earth Sciences. During the MSc project professor Fokkema showed me the good things of
seismic research. At the end of the academic year, August 1991, professor Fokkema
introduced me to the group of professor Berkhout where I could begin a PhD study
within the DELPHI project. Research in the Geophysical context was something I
started to like a lot and working in a team was also an important reason for starting
my PhD project in the group of professor Berkhout. The subject of my PhD project
was initially set up to investigate the combination of elastic wavefield decomposition and weathered layer influences. In the first year it became clear that correct
estimation of the propagation properties of the weathered layer is very important
for the decomposition and all the further processing steps. Therefore the subject
of my research changed to weathered layer estimation, using the areal shot record
technology. An important side-off project of these first years was an optimization
technique I used to construct the decomposition operators. This technique is also
very successful for the optimization of extrapolation operators. During my work
with areal shot records the areal shot record designed for one point in the subsurface turned out to be a very useful intermediate step for imaging and velocity
analysis. So finally the subject changed for the last time to Common Focus Point
technology the main subject of this thesis. The last 2 years I have been working
intensively on this subject and the results can be found back in this thesis. Note
that although it seems that I have been doing a lot of other things besides the main
subject of this thesis, all the things I have done before helped me to understand the
main geophysical problems and gave me a broad overview.
During the past 5 years I have learned a lot from professor Berkhout, Kees Wapenaar
and Eric Verschuur, who are the driving forces behind the DELPHI research team.
The discussions I have had with my fellow PhD students were always stimulating and
very useful for understanding the practical implications of the theory. The DELPHI
consortium, which is sponsored by the oil and computer industry, reports twice a
year the research results at the so called sponsor-meetings. Once a year a book,
vi
Preface
with the latest scientific results, is published for the sponsors. At first I found it
difficult to give oral presentations at the sponsor meetings, but during the years I
have learned how to give a presentation and in the last year I hope the sponsors
could follow what I was saying. At this point I wish to thank the participating
companies for making the research within DELPHI possible, and for their interest
and comments they gave at the sponsor meetings.
There are many people who have helped me in the last 5 years. First of all I would like
to thank my promoter, professor Berkhout, for supervising this thesis, his enthusiasm
about the subject and his stimulating ideas. The comments and suggestions Kees
Wapenaar gave me during the years, especially for the theoretical part of this thesis,
have been very useful and gave the theoretic parts its clear structure. Eric Verschuur
always knew the answers to all my questions, and independent of the subject the
answer was always useful.
Beside all the new knowledge about geophysics I’ve also learned how to work in a
team. My colleagues of the first year Greg, Cees and Erwin helped me starting my
research. Walter learned me everything, I always wanted to know, about areal shot
record technology and showed me how to make up reasons for not running during
lunch-time. I also would like to thank Aart-Jan, Frederic and all the incidental
’runners’ for running with me during lunch-time. I’m thankful for the patience of
Nurul, and all the other students who have worked with my programs, with the
build in features/bugs of my programs. During the years Riaz became a very good
friend and we have had a lot of good times together. The boys of next door Felix,
Frank and Wim are remembered for being very quiet neighbors. I will miss Felix for
letting me know the latest news on the wave-equation, and Frank for providing the
running group with sun-milk during the run.
Without the system management of JWdB, Leen, Edo and Henry a lot of work
couldn’t be done at all. JWdB and Felix introduced NEXTSTEP into the group,
which resulted in an important improvement in the development of applications
and state of the art pictures. Although Alexander is not a part of the system
management I will never forget his famous Unix scripts and his willingness to help
to get the system up again, after a serious crash. I also would like to thank Bart
and Scott of Cray Research in Eagan for offering me a great job at Cray Research /
Silicon Graphics. I wish the new PhD students the same good time I have had in the
DELPHI team, specially the never ending good weeks during the sponsor meetings
and geophysical congresses.
The last months, while I was writing this thesis, my family and Roos kept me on
the right track and helped me remembering that there are more things to life than
writing a thesis.
Contents
Preface
v
Notation and Terminology
1 Introduction
ix
1
1.1
The importance of imaging . . . . . . . . . . . . . . . . . . . . . . .
1
1.2
The forward model . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3
1.3
Outline of this thesis . . . . . . . . . . . . . . . . . . . . . . . . . . .
4
2 Migration: an overview
5
2.1
Isochrone summation . . . . . . . . . . . . . . . . . . . . . . . . . . .
5
2.2
Finite difference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7
2.3
Kirchhoff summation . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
2.4
Migration in terms of deconvolution . . . . . . . . . . . . . . . . . .
13
2.5
Inverse scattering . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15
2.6
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
16
3 Two-way and one-way representations
3.1
19
Reciprocity theorems . . . . . . . . . . . . . . . . . . . . . . . . . . .
19
3.1.1
Reciprocity theorem for two-way wavefields . . . . . . . . . .
20
3.1.2
Reciprocity theorem for one-way wavefields . . . . . . . . . .
21
3.2
Integral representations for two-way wavefields . . . . . . . . . . . .
25
3.3
Integral representations for one-way wavefields . . . . . . . . . . . .
28
3.4
The WRW model in matrix notation . . . . . . . . . . . . . . . . .
35
viii
Contents
4 Imaging by double focusing
39
4.1
Inverse scattering problem . . . . . . . . . . . . . . . . . . . . . . . .
39
4.2
Two-way representation of double focusing; full aperture . . . . . . .
41
4.3
Two-way representation of double focusing; seismic aperture . . . . .
44
4.4
One-way representation of double focusing . . . . . . . . . . . . . . .
45
4.4.1
Integral representation . . . . . . . . . . . . . . . . . . . . . .
46
4.4.2
Matrix representation . . . . . . . . . . . . . . . . . . . . . .
49
5 CFP technology
53
5.1
Areal shot record technology . . . . . . . . . . . . . . . . . . . . . .
53
5.2
First focusing step . . . . . . . . . . . . . . . . . . . . . . . . . . . .
55
5.2.1
Huygens-Fresnel principle . . . . . . . . . . . . . . . . . . . .
57
5.2.2
Construction of CFP trace: dipping layer . . . . . . . . . . .
59
5.2.3
Construction of CFP gathers: synclinal model . . . . . . . . .
65
5.2.4
Focusing in emission and focusing in detection . . . . . . . .
70
Second focusing step . . . . . . . . . . . . . . . . . . . . . . . . . . .
71
5.3.1
One-way image ray . . . . . . . . . . . . . . . . . . . . . . . .
73
5.3.2
Positioning of the double focusing result . . . . . . . . . . . .
74
5.3.3
CFP image-gather . . . . . . . . . . . . . . . . . . . . . . . .
78
5.3.4
One-way time image . . . . . . . . . . . . . . . . . . . . . . .
79
Resolution and amplitude analysis . . . . . . . . . . . . . . . . . . .
83
5.4.1
Resolution and focusing beams . . . . . . . . . . . . . . . . .
83
5.4.2
Amplitude analysis . . . . . . . . . . . . . . . . . . . . . . . .
87
3-Dimensional CFP gathers . . . . . . . . . . . . . . . . . . . . . . .
91
5.5.1
Common offset contributions in 2-dimensional synthesis . . .
91
5.5.2
A simple 3D data example . . . . . . . . . . . . . . . . . . . .
94
5.5.3
Regularization of coarsely sampled data . . . . . . . . . . . . 100
5.3
5.4
5.5
5.6
New developments . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
6 Operator updating
105
6.1
First focusing step . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
6.2
Searching for updating formulas for a flat reflector . . . . . . . . . . 111
6.2.1
Zero depth and zero velocity error (∆z = 0 and ∆c = 0) . . . 113
Contents
6.3
6.4
ix
6.2.2
Depth errors (∆z 6= 0 and ∆c = 0) . . . . . . . . . . . . . . . 114
6.2.3
Velocity errors (∆z = 0 and ∆c 6= 0) . . . . . . . . . . . . . . 115
6.2.4
CFP corrected shot record . . . . . . . . . . . . . . . . . . . . 115
6.2.5
Move-out corrected CFP gather . . . . . . . . . . . . . . . . . 117
Operator updating . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
6.3.1
Flat reflector . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
6.3.2
Dipping reflector . . . . . . . . . . . . . . . . . . . . . . . . . 127
Second focusing step . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
7 Numerical data examples
137
7.1
Diffraction point . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
7.2
One dimensional multi layer model . . . . . . . . . . . . . . . . . . . 140
7.3
’Void’ model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
7.4
Weathered layer model . . . . . . . . . . . . . . . . . . . . . . . . . . 148
7.5
Syncline model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
7.6
Comparison of imaging results for realistic numerical data . . . . . . 154
7.6.1
Marmousi model . . . . . . . . . . . . . . . . . . . . . . . . . 154
7.6.2
SEG/EAGE salt dome model . . . . . . . . . . . . . . . . . . 162
8 Field data examples
167
8.1
Mobil . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
8.2
ELF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
A Operator optimization
185
A.1 1-Dimensional operators for 2-dimensional extrapolation . . . . . . . 185
A.1.1 Analytical space-frequency operators . . . . . . . . . . . . . . 186
A.1.2 From wavenumber domain to spatial convolution operators . 188
A.1.3 Recursive depth migration . . . . . . . . . . . . . . . . . . . . 198
A.2 2-Dimensional operators for 3-dimensional extrapolation . . . . . . . 205
A.2.1 Direct method . . . . . . . . . . . . . . . . . . . . . . . . . . 207
A.2.2 McClellan transformation, expansion in cos (kr ) . . . . . . . . 227
A.2.3 Expansion in kz
A.2.4 Expansion in
kx2
. . . . . . . . . . . . . . . . . . . . . . . . . 242
+ ky2
. . . . . . . . . . . . . . . . . . . . . . 250
x
Contents
A.2.5 Computation times . . . . . . . . . . . . . . . . . . . . . . . . 256
A.2.6 Concluding remarks . . . . . . . . . . . . . . . . . . . . . . . 258
B Matrix notation
263
B.1 2-Dimensional wavefields . . . . . . . . . . . . . . . . . . . . . . . . . 263
B.2 3-Dimensional wavefields . . . . . . . . . . . . . . . . . . . . . . . . . 266
C Algorithms
269
C.1 Processing flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
C.2 Time- and Frequency-domain processing . . . . . . . . . . . . . . . . 270
C.3 Numerical implementation (time-domain) . . . . . . . . . . . . . . . 272
Bibliography
277
Summary
287
Samenvatting
289
Curriculum vitae
291
Notation and terminology
The subscript notation for Cartesian vectors and tensors is used. Lower case
Latin subscripts {k, l, p, q} are assigned to the values 1, 2 and 3 and lower
case Greek subscripts {α, β} are assigned to the values 1 and 2. The summaP3
tion convention applies to repeated subscripts (e.g. ak bk stands for
k=1 ak bk ).
Differentiation of a function with respect to time is denoted as ∂t hk , which
stands for {∂t h1 , ∂t h2 , ∂t h3 }. Differentiation of a vector with respect to the
spatial coordinates is denoted as ∂k hl and is the shorthand notation for
{{∂1 h1 , ∂1 h2 , ∂1 h3 }, {∂2 h1 , ∂2 h2 , ∂2 h3 }, {∂3 h1 , ∂3 h2 , ∂3 h3 }}. The gradient of a scalar
function is written as ∂k φ which stands for {∂1 φ, ∂2 φ, ∂3 φ}, the divergence of a vector function is notated as ∂k hk . To register a position in space a Cartesian reference
frame with an origin O and three mutually perpendicular base vectors {i1 , i2 , i3 } of
unit length is used.
Functions in the space-time domain are denoted by a lower case symbol, f (x, t). The
corresponding function in the space-frequency domain are denoted by a upper case
symbol F (x, ω). Functions in the wavenumber-frequency domain have a tilde above
their symbol F̃ (kα , ω). Matrices are denoted in the space-time domain by a bold
lower case symbol p, in the space-frequency domain by a bold upper case symbol
P and in the wavenumber frequency domain a tilde above the symbol is used P̃.
Vectors are in bold italic fashion p, P , P̃ . For discrete matrices and discrete vectors
sans-serif fonts are used. These notation conventions are summarized in the table
given below. Note that there is no difference between an operator in the time-domain
and an operator in the frequency domain. However, by looking at the function where
the operator is working on, the domain can be derived.
With the above definitions an operator L working on the function u can be represented by the convolution integral
L(x)u(x) =
Z
x1
L(x, x′ )u(x′ )dx′ .
x0
In discrete notation the integral is replaced by a summation and can be written in
xii
Notations and Terminology
symbol
xk − t
xk − ω
kα , x3 − ω
function
f (xk , t)
F (xk , ω)
F̃ (kα , x3 , ω)
vector
p
P
P̃
matrix
p
P
P̃
operator
L
L
L̃
discrete vector
p
P
P̃
discrete matrix
p
P
P̃
discrete operator
L
L
L̃
a matrix vector multiplication given by



u0
L0,0 . . . L0,M
  .. 

..
..
Lu =  ...
 . .
.
.
uM
LN,0 . . . LN,M
Note that matrices which represent discrete operators or wave fields are written
with square [ ] brackets. Matrices with continuous operator kernels are written with
normal ( ) brackets
The temporal Fourier transformation from the space-time domain to the spacefrequency domain is defined as (Bracewell, 1986)
Z ∞
F (x, ω) =
f (x, t) exp (−jωt)dt
−∞
and its inverse as
f (x, t) =
1
ℜ
π
Z
∞
0
F (x, ω) exp (jωt)dω
where ℜ stands for the Real part. The 2-dimensional spatial Fourier transformation
from the space-frequency to the wavenumber-frequency domain is defined as
F̃ (kα , ω) =
+∞
ZZ
F (xα , ω) exp (jkα xα )d2 xα
−∞
and its inverse as
F (xα , ω) =
1
2π
2 +∞
ZZ
−∞
F̃ (kα , ω) exp (−jkα xα )d2 kα
Notation and Terminology
xiii
The transpose of a matrix (or vector) is denoted with a superscript T , the complex
conjugate with a superscript ∗ and the complex conjugate transpose with superscript
H so

 ∗
P1
P1
 P2 
P2∗ 
 
 
P =  .  P∗ =  . 
 .. 
 .. 

Pn
Pn∗
P T = (P1 P2 . . . Pn )
P H = (P1∗ P2∗ . . . Pn )
Terminology
The new concepts used in the CFP approach has led to a number of new terms
which are explained below.
• focusing operator
Convolution operator working in the time domain on pre-stack seismic data.
The operator represents the response of a point in the subsurface measured at
the surface and works on the traces in a common detector gather or a common
shot gather. Summation over the resulting traces in the gather defines one
trace of a CFP gather.
• focusing in detection
Result of the focusing operator working on a common shot gather. The result
can be interpreted as the measurement of an areal receiver positioned at the
focus point in the subsurface, which is related to the focusing operator, and
the source position of the used common shot gather.
• focusing in emission
Result of the focusing operator working on a common receiver gather. The
result can be interpreted as the response of an areal source positioned at the
focus point in the subsurface, which is related to the focusing operator, and
the receiver position of the used common receiver gather.
• common focus point (CFP) gather
The multi-offset response of the subsurface due to a focusing areal source. For
focusing in detection the measurements are generated by individual sources
with different positions at the surface. For focusing in emission the response
is registered by individual receivers with different positions at the surface.
xiv
Notations and Terminology
• focus point response
The coherent event in a CFP gather, that represents the reflection response
from the involved focus point. For a correct macro model the traveltimes
defined by the operator and the focus point response are equal (’principle of
equal traveltime’).
• one-way offset
The lateral distance between the position of the focus point and the position
of a trace in the CFP gather.
• one-way image time
The time defined by the one-way zero-offset trace in the focusing operator.
• one-way move-out
Difference in traveltime between the one-way image time and the times given
by the focus point response.
• CFP image condition
The imaging condition in the CFP method states that a reflection area exists
within the earth when the traveltime of the wavefront of the downward continued wavefield is the same as the traveltime of the wavefront of the downward
incident wavefield for all offsets.
• one-way image gather or CFP image gather
The one-way image gather is constructed of time windows selected from moveout corrected CFP gathers. The CFP gathers used to construct the image
gather have their focus points defined at the same lateral position in the model,
but at different depth (or time) levels. The lateral position of the image gather
is the same as the position of the focus points used to construct the image.
Note that the image trace is constructed by a summation over the traces in
the image gather an is defined in one-way image time or in depth.
• CFP corrected shot record
The CFP corrected shot record is the result of the common shot gather after
temporal convolution with the focusing operator but before summation over
the traces in the gather. In this corrected shot record the Fresnel zone can be
observed.
Notation and Terminology
xv
Table of notation
{i1 , i2 , i3 } = the basis vectors in 3-dimensional Euclidean space
{x1 , x2 , x3 } = {x, y, z} = orthogonal Cartesian coordinates
x = x1 i1 + x2 i2 + x3 i3 = vectorial position
D = some bounded domain in 3-dimensional Euclidean space
3
d x = elementary volume in 3-dimensional Euclidean space
∂D = the boundary of D
d2 x = elementary area of ∂D
λ = wavelength [m]
δ(x) = spatial delta function
H(x3 ) = Heaviside step function
1
χ(x) = 1, , 0 when x ∈ D, ∂D, R3 \(D ∪ ∂D) ; characteristic function
2
a◦ = indicating a dip of ’a’ degrees
P(zr , zs ) = matrix representation of seismic data (see Appendix C)
zr = receiver level, in general a function of xk
zs = source level, in general a function of xk
ρ = the volume density of mass [kg/m3 ]
κ = the compressibility [Pa−1 = m2 /N]
W+ (zm , z0 ) = extrapolation of downgoing wavefields from z0 to zm
W− (z0 , zm ) = extrapolation of upgoing wavefields from zm to z0
F+ (z0 , zm ) = inverse extrapolation of downgoing wavefields from zm to z0
−1 −
∗
= W+ (zm , z0 )
≈ W (z0 , zm )
F− (zm , z0 ) = inverse extrapolation of upgoing wavefields from z0 to zm
−1 +
∗
= W− (z0 , zm )
≈ W (zm , z0 )
T
W+ (zm , z0 ) = W− (z0 , zm )
T
F− (zm , z0 ) = F+ (z0 , zm )
R+ (zm ) = matrix representation of reflection operator
Fi− (zm , zr ) = operator for focusing in detection at focal point x = {(x, y)i , zm }
Fj+ (zs , zm ) = operator for focusing in emission at focal point x = {(x, y)j , zm }
Pi− (zm , zs ) = CFP gather for focusing in detection
Pj (zr , zm ) = CFP gather for focusing in emission
xvi
Notations and Terminology
Chapter 1
Introduction
The seismic method is based on measurements from detectors placed at the surface,
or in the subsurface, of the earth. The detectors measure the wavefield originating
from a source which is positioned at or close to the surface. In figure 1.1 three
types of seismic experiments are shown: a marine, land and borehole. In the marine
and land configuration a part of the energy emitted by the source travels along the
surface of the earth directly to the receivers. This wave is called the direct or surface
wave. The surface wave contains information about the layers close to the surface
of the earth. However, the aim in seismic processing lies in the detection of the
structure of the deeper layers (subsurface). Besides the surface wave the source also
transmits waves into the subsurface, called the body waves. Due to contrasts in
the subsurface a downward propagating body wave gets reflected and propagates
back towards the surface where it can be measured by the detectors. These reflected
wavefields contain the information the geophysicist is interested in. The goal of the
geophysicist is to derive from these reflections an accurate image of the subsurface
of the earth. The technique to translate the measurements at the surface (called
shot records) into a structural representation of the subsurface is called imaging or
migration. For a good structural image a large number of shot records is needed,
where for every shot record the source and receivers are placed at another position
at the surface.
1.1
The importance of imaging
In the left-hand side picture of figure 1.2a a raw shot record is shown. This shot
record contains the surface related waves and the body waves. For the construction
of an image the surface waves are not directly needed and can be removed from
the raw shot record. Besides the primary reflections of the layer boundaries, the
shot record also contains reflections from the surface of the earth. These waves
are called surface related multiples and have travelled more than one time through
2
1.1 The importance of imaging
geophone
source
hydrophone
airgun
∆ ∆∆∆ ∆ ∆ ∆∆∆∆ ∆
well log
Figure 1.1
Three types of seismic data acquisition; (1) acquisition at sea where hydrophones are
used to measure the response (pressure) from an airgun array near the surface, (2) acquisition on land where geophones are used to measure the response (particle velocity in
3-directions) from a source in the surface, (3) acquisition in a borehole where geophones
or hydrophones are mounted on a cable (which is lowered in the borehole) to measure
the response from a source at the surface.
the subsurface of the earth. These surface related multiples can distort the image
quality significantly and must be removed from the data. After these pre-processing
steps the resulting shot record is shown in figure 1.2. From this figure the primary
reflections from the subsurface are better visible. However, it is still not clear where
in the subsurface these reflections are actually originating from. By making use of an
imaging technique the time traces of all pre-processed shot records are transformed
into depth traces. For this process a macro model of the earth is needed. This macro
model describes the propagation properties of the earth, in particular with respect to
traveltimes. After the imaging step a geologist can interpret the result and identify
the structures and layers in the subsurface of the earth. This interpretation can, for
example, be used to make a decision about the position of a future borehole.
If an error has been made in the imaging procedure the structural image, and the
interpretation based on it, will be wrong and the borehole may miss its target. Therefore, an error analysis of the quality of the image is very important. Unfortunately,
this error analysis can only be carried out successfully if one knows the correct answer, which is equal to the goal of seismic imaging. In this thesis a novel imaging
technique is proposed in which the error analysis is carried out in an intermediate
domain where the operators, used to construct the image, are compared with the
’half image’. Based on this comparison the macro model of the earth or, even better,
the operators themselves can be updated. The proposed imaging technique makes
use of a double focusing procedure in which the result after one focusing step, the so
Chapter 1: Introduction
0
-3000
offset [m]
-2000
-1000
0
0
offset [m]
-2000
-1000
0
1
time [s]
time [s]
1
-3000
3
2
2
3
3
4
4
Figure 1.2
Seismic shot record for a marine acquisition before pre-processing (left) and after preprocessing (right). In the pre-processing step the direct waves and the multiple reflections from the sea surface are removed.
called Common Focus Point (CFP) gather, plays an important role. The theoretical
frame work, where the proposed imaging technique is developed in, is the systems
oriented WRW model (Berkhout, 1982).
1.2
The forward model
To describe the seismic experiment in mathematical and physical terms a forward
model is needed. The forward model used in this thesis is shown in figure 1.3.
The top part of figure 1.3 shows the model used in the imaging, the bottom box of
figure 1.3 represents the lithologic inversion. In imaging the aim is to estimate the
reflectivity operator R by removing the propagation parts W+ , W− and the surface
related effects D+ , D− . In lithologic inversion the aim is to estimate the rock type
e.g. a sand or a shale layer. The connection between the imaging model and the
lithological model is the angle dependent reflection property of the layer. In the
imaging technique proposed in this thesis the reflection property can be determined
in a straightforward way. The different boxes shown in the forward model for the
imaging part will become clear in the remainder of this thesis. For the lithologic
inversion part the reader is referred the thesis of de Bruin (1992).
4
1.3 Outline of this thesis
P(zr , zs )
D− (zr )
R− (zs , zr )
+
D+ (zs )
S0 (zs )
W+ (zm , zs )
W− (zr , zm )
R+ (zm )
lithologic
model
Figure 1.3
1.3
Full WRW forward modeling scheme for seismic data. The connection with the lithologic model is made by the reflection matrix.
Outline of this thesis
Chapter 2 of the thesis starts with a historical overview of seismic imaging to provide
a historical context of the seismic imaging technique presented. The forward model
used to derive the double focusing procedure is explained in chapter 3. Starting from
the one-way and two-way acoustic reciprocity relations integral and matrix representations of seismic data are derived from which the forward model is constructed.
Based on the same representations the double focusing procedure is introduced in
chapter 4 and explained for one-way and two-way wavefields. At the end of chapter
4 the matrix notation of the double focusing technique is presented. The theory
presented in chapter 3 and 4 is illustrated with numerical experiments in chapter
5. For those readers who want to skip the theoretical chapters, chapter 5 is a good
starting point to get an understanding of the possibilities of the Common Focus
Point technology. The influence of erroneous focusing operators and the updating
of the operators is explained with simple numerical examples in chapter 6. Results
on numerical modeled data can be found in chapter 7 and results for field data can
be found in chapter 8. The numerical data examples are used to show the strength
and weakness of the double focusing method. The results obtained with field data
are compared with results obtained with other imaging methods.
Appendix A gives an extensive overview of methods to construct extrapolation operators (2D and 3D) for explicit recursive depth extrapolation. The resulting operators
are used in an extrapolation algorithm to construct focusing operators in complex
subsurfaces. Appendix B explains the matrix notation used in this thesis and appendix C discusses the numerical schemes of the CFP technology. On page ix an
overview is given of the notation conventions and definitions used in this thesis. At
the end of the thesis a summary is given with the most important conclusions
Chapter 2
Migration: an overview
The purpose of this chapter is to provide a historical context of the seismic imaging
(migration) technique. It is impossible to give a complete overview of the seismic
migration theory in this chapter. Hence, only those parts of the migration history
are discussed which will contribute to the understanding of the theory presented.
In this chapter the main periods of the migration history are briefly discussed: in
a first period the concepts of migration are developed by making use of graphical
methods, in the second important period the idea of reflector mapping is introduced
and in the last main period computation intensive wave equation based methods
for 3-dimensional data are used for imaging1 . This chapter does not include an
overview of the wide range of migration based velocity or macro model estimation
techniques. This absence does not mean that the importance of velocity estimation
is underestimated; without a good velocity model every depth imaging method will
break down.
Migration is the technique used to transform the wavefield of a (stacked) seismic
section into a reflectivity image. The migration process influences the position of the
reflectors as well as the resolution property along the reflectors. For most migration
methods the wave equation is assumed to be the mathematical description of wave
propagation in the earth’s subsurface.
2.1
Isochrone summation
The earliest migration techniques were graphical and based on geometrical ideas
developed systematically by Hagedoorn (1954). Hagedoorn describes migration in
terms of propagating wavefronts and tries to avoid the use of non-physical ray paths.
Hagedoorn argues that within the seismic frequency band it is impossible to speak
of rays in a physical sense of narrow beams. The principle of Huygens-Fresnel,
1 For a complete overview of migration up to 1985 the collection of reprints of Gardner (1985)
is recommended.
6
2.1 Isochrone summation
Ts
Figure 2.1
xs
xr
Tr
A single source receiver combination defines for every reflection time a surface of equal
traveltimes. The vertically plotted point at the surface of equal traveltimes in the middle
between source and receiver is used to determine a surface of equal reflection times.
explained in more detail in chapter 5, shows that the beam between source and
receiver is at least a half wavelength wide. Therefore it is conceptually better to
work with propagating wavefronts than with rays; the ray can be considered as a
mathematical abstraction defined as the line perpendicular to the wavefront at each
intermediate time. Seismic waves exhibit real diffraction phenomena such as bending
around obstacles as well as focusing associated with transmission and reflection
from curved interfaces. These phenomena cannot be described correctly by making
use of ray-paths. According to Hagedoorn the aim of migration is to position the
reflected energy from the wavefront measured at the surface at its correct position
in the subsurface. From one reflection arrival time, belonging to a source receiver
combination, a so called surface of equal reflection time can be constructed. In figure
2.1 a set of wavefronts, centered at the source position xs and the receiver position
xr , is shown. A reflection time of 2T [s] observed at xr can originate from any point
on the surface (in a 3-dimensional sense) of equal traveltime consisting of lines of
intersection between wavefront surfaces T+t [s] from xs with wavefront surfaces T-t
[s] from xr , which is indicated by the thick line in figure 2.1.
So the position of the reflector is not yet known from one single observation, but
the surface of equal reflection times is known to be tangential to the actual reflector
at some point in space. It is convenient to represent the surface of equal reflection
time by one point on it, normally the Common Mid Point (CMP) between source
and receiver is chosen as ’reference’ point to describe the surface of equal reflection
time. This vertically plotted point has no other significance than that of being one
point determining a surface of equal reflection times.
Hagedoorn (1954) defined migration as ”the procedure of determining the true re-
Chapter 2: Migration: an overview
7
P (migrated)
curve of equal
reflection times
Figure 2.2
Q (plotted vertically)
curve of
maximum
convexity
2-Dimensional graphical procedure to obtain from the curve of equal reflection times and
the curve of maximum convexity the migrated position (P). The dotted line represents the
vertically plotted horizon through Q.
flecting surface from a surface determined by a number of vertically plotted points”.
This true surface can be found, in principle, as the envelope to all surfaces of equal
reflection time determined by the vertically plotted points of different source receiver combinations. The apparent horizon of several vertically plotted points form
a surface of maximum convexity for the reflector. By using these two surfaces, the
vertically plotted points and the surface of equal traveltimes, the migration can be
carried out in a graphical manner as shown in figure 2.2. In figure 2.2 a chart of
curves of equal reflection times is centered on the vertical around Q and the chart of
curves of maximum convexity is moved until the best tangential fit to the vertically
plotted horizon at Q is obtained. Both the traveltime curves and the maximum
convexity curves are based on a (1-dimensional) background model, For any other
velocity distribution a new set of charts should be calculated. This method of Hagedoorn (1954) is a graphical procedure where the computer was used to generate the
wavefront charts. The surface of maximum convexity (also called Huygens surface)
is the ’kinematic image’ in the time domain of a point in the depth domain. The
surface of equal reflection time (isochrone), on the other hand, is the ’kinematic
image’ in the depth domain of a point in the time domain (Hubral et al., 1996).
2.2
Finite difference
Given waves observed along the earth’s surface it is possible, by using some mathematical techniques, to extrapolate these waves into the earth. In this approach the
migration process progressively transforms the wavefield measured at the earth’s
surface into wavefield’s that would be observed at progressively increasing depths.
This so called inverse extrapolation technique transforms the measured wavefield to
virtual receivers on a depth closer to the reflector(s). The basic idea behind it is that
the best measurement of any reflector is when the receivers are placed just above the
reflector. These extrapolated wavefields are not yet the migrated section, therefore
8
2.2 Finite difference
next time
level
stacked
data
Figure 2.3
extrapolation
imaging
migrated
time
section
Migration can be considered as the results of two steps; extrapolation and imaging.
an additional imaging step is needed. The imaging principle most seismic imaging
methods use are based on the basic principle of reflector mapping introduced by
Claerbout (1971). By backpropagating the scattered field from the (combination
of) receiver array(s) into the background medium the reflected wavefield is reconstructed in the medium. To image this backpropagated field at every point in the
medium the extrapolated field is correlated with the forward extrapolated incident
wavefield. This incident wavefield is the solution of the forward problem in the
background medium and is obtained by placing a (combination of) source(s) at the
surface. If the background model (macro model) is correct then at every reflector
the upward reflected field should be equal to the downward source wavefield multiplied with the reflection coefficient. A correlation between the reflected field and
the incident field images the reflector at zero-time. So ”reflectors exist at points in
the earth where the first arrival of the downgoing wave is time coincident with an
upgoing wave”. In the method of Claerbout implicit use is made of one-way wave
propagation through inhomogeneous media. In this one-way wave method only the
transmitted field is assumed to be of importance, by assuming that the earth is only
weakly inhomogeneous and therefore only a small fraction of the total energy applied at the surface returns to it. The internal multiple reflections are also neglected
resulting in an image result with less interference effects from these multiples (see
also Berkhout and Wapenaar, 1989).
In this approach seismic migration is considered as an acoustic image reconstruction
technique which makes use of two steps as shown in figure 2.3;
• wavefield extrapolation, to simulate registrations at other depth levels
• imaging, to image the extrapolated wavefield
To calculate the incident wavefield and the backpropagated reflected wavefield at
all different depth levels of interest several wavefield extrapolation techniques have
been introduced. The finite difference method, introduced by Claerbout (1971),
uses a discretized version of the wave equation with a floating time reference. The
disadvantages of the proposed finite difference implementations are the time domain
approach (’time migration”) and the small extrapolation angles (typically 15 − 45◦).
Chapter 2: Migration: an overview
9
In the well-known stacking procedure the traces in the shot records are combined to
Common Mid Point (CMP) sections and corrected for the source-receiver geometrical
effects (called the Normal Move Out (NMO) correction). Note that the described
stacking procedure means that the vertically plotted points of Hagedoorn (1954)
are combined into one trace for several source receiver combinations with the same
mid point. The process of stacking is also used to provide a good estimate (from a
signal to noise point of view) of what a coincident source receiver pair would record
(zero-offset). For a 1-dimensional medium, meaning that there are only medium
changes in the depth (z) direction and not in the lateral directions (x and y), the
’migration’ of the move-out corrected and stacked section consists of only a stretch
in the time direction to map the time axis to the depth axis. For laterally changing
media it is necessary to employ migration techniques to focus the data.
2.3
Kirchhoff summation
The Kirchhoff summation method has its basis in ray path and traveltime considerations and the diffraction theory of Huygens and Fresnel. French (1974) describes
migration as a process in which each subsurface point is assigned a number which is
a measure of the probability that scattered energy has emanated from that possible
scattering area. The number is determined by summing the recorded data for all
shot points and receiver locations at times where energy from that subsurface point
could arrive. This is equivalent to a summation along the hyperbolic Huygens surface (surface of maximum convexity). This operation is repeated for every sample
on the seismic output section. The hyperbolic traveltime curves, which define the
path of the integration, used in this method are calculated by making use of stacking velocities. The advantage of this technique is that it is possible to migrate 3D
seismic data within a reasonable computation time. The disadvantage is that the
method is based on kinematics only and the wave equation is not used explicitly.
The diffraction summation method implies that the geologic interface acts as a diffuse
reflector; that is, each point is assumed to scatter energy all along its corresponding
reflection-time surface (French, 1975). So a wavefield at the earth’s surface (a seismic
section) is interpreted as a superposition of an infinitude of smaller fields from the
distribution of scatterers. A reflection events is thus the outcome of constructive
and destructive interference of an infinitude of infinitely weak diffraction patterns.
In figure 2.4 the principle used in the diffraction summation method is shown for
a dipping reflector. The reflector is build by making use of a limited number (in
figure 2.4 16 diffraction points are used) of diffraction points positioned on the
reflector. The apexes of the diffraction hyperbola define the correct position of the
reflecting events, which is indicated with the thick line in figure 2.4. Constructive
interference builds up a reflector along the straight-line envelope of the diffraction
curves. According to this reasoning the goal in migration is to transform these
10
2.3 Kirchhoff summation
-1500
0
-750
lateral position [m]
0
750
1500
time [s]
0.5
1.0
1.5
2.0
Figure 2.4
Diffraction stack of a dipping reflector in a homogeneous medium (c = 2000 [m/s]). The
thick line, defined through the apexes of the individual diffraction responses, represents
the true position of the dipping reflector.
diffractions, and the reflection build up from it, into a reflector segment given by
the thick line. The diffraction summation approach exploits this relation between
diffraction pattern and scatterer location. A possible subsurface scatterer will have
a nearly hyperbolic diffraction curve (assuming a 1-dimensional medium) in the
unmigrated time section with an apex at the sample point of the scatterer. Migration
then involves summation of the input amplitudes along the diffraction curve and
placing the sum at the output apex position. This operation is repeated for every
sample point on the seismic output section.
The diffraction summation method approximates the propagation of waves in the
earth, it does not describe a physical event. In order to account for frequencydependent amplitude and phase behavior, wave propagation theory must be introduced in the method. Therefore a new, more physical related, interpretation of
the stacked seismic section was introduced by Loewenthal et al. (1976). This interpretation became known as the exploding reflector model. At each reflector in
the subsurface sources are placed with charges having a local strength proportional
to the effective reflectivity. At time t = 0 all sources are fired simultaneously and
only the upward traveling waves are propagating to the surface through a medium
with a velocity halve of the true velocity. The resulting wavefield at the earth’s
surface approximates a normal-incidence (with respect to the surface) time section.
Loewenthal et al. (1976) used this interpretation to migrate a stacked seismic section with the finite difference method of Claerbout (1971). Within this model it
is assumed that the ray paths between a source-receiver location and a point on
the reflector is the same for upward and downward propagation (representing the
Chapter 2: Migration: an overview
11
wave
equation
drop some
terms
Kirchhoff
integral
new
equation
ignore
some terms
finite
difference
summation
imaging
principle
image
Figure 2.5
Two wave equation based approaches to migration (after Larner and Hatton, 1990);
Kirchoff migration and finite difference.
normal ray path). In the exploding reflector model the assumption is made that
stacked data is equivalent with zero source-receiver offset. Within this interpretation the wavefield observed in the stacked section can be extrapolated downward
into the subsurface by using wave equation based methods followed by an imaging
principle to image the reflectors (French, 1975). However, using the WRW model,
Berkhout (1982) derived the theoretical foundation of the exploding reflector model
and showed the conditions for the validity of such a model by transforming WRW
to a W0 R model for zero-offset data.
Schneider (1978) showed that the Kirchhoff integral formulation has strong ties
with the diffraction summation approach, the differences are subtle but significant
in terms of amplitude and waveform reconstruction. The Kirchhoff integral method
has no limitations on the reflector dip and in principle angles > 90◦ can be migrated.
Schneider (1978) also showed that the Kirchhoff integral formulation of the diffraction summation approach for stacked seismic data can be used to simulate the angle
limitations of finite difference algorithms.
For wave equation based migration the seismic section is posed as a boundary value
12
2.3 Kirchhoff summation
problem for the wave equation. Solutions can be derived from either an integral
or differential form of the wave equation (see figure 2.5). The method of summation along diffraction hyperbolas is founded on the integral solution. The migration
scheme of Claerbout (1971) is based on a solution of differential form (Larner and
Hatton, 1990). Both the integral (summation) approach and the differential approach yield approximate solutions of the same wave equation. However, specific
approximations inherent in actual computational algorithms differ for the two approaches. Note that both approaches are forms of time migration; so a macro model
which described the propagation properties of the subsurface as function of depth is
not used explicitly. This means that due to propagation through laterally varying
media closer to the surface the shape of the two-way time image of deeper reflectors
is altered. Therefore an additional depth migration, which shift time migrated points
laterally and vertically to their correct position in depth, is required for proper imaging of the time migrated section when the subsurface overburden exhibits lateral
complexity. In Black and Brzostowski (1994) these kind of errors made with time
migration are clearly explained.
The Kirchhoff summation method sums signals into the apex of the approximation
hyperbolic (Hubral, 1977). The position of the time surface minimum (apex) is the
position where a ray from a depth point emerges vertically to the surface. Thus
time migration of data has the effect of moving points laterally to their minimum
time position (image ray) rather than their true lateral position. This means that a
Kirchhoff migration must be followed by an additional time to depth migration for
the true depth position to be recovered. This deficiency of Kirchhoff time migration
is sometimes called the ’lack of Snell’s law’, which means that the breaking of the
ray along interfaces is not taken into account; the image ray is interpreted as it
has vertically traveled through the layers of the earth without breaking. The image
ray is defined as the minimum traveltime path from a point on the reflector to the
surface and will always emerge vertically at the surface. Note that the image ray is
associated with time migration and not with the true diffraction curve. Related to
the definition of the image ray is the normal ray which is defined as the ray which
travels along the minimum time path from the surface to a reflector and ends, by
definition, perpendicular to the target interface (Parkes and Hatton, 1987). Figure
2.6 shows both the image and the normal ray.
The signal positions and two-way times related to the image-rays are not affected
by the time migration process, signals for all other rays are migrated. A zero-offset
trace (a trace out of a stacked section is assumed to represent a zero-offset trace) is
transformed by the migration process into an image ray belonging to a point close to
the normal incidence point of the normal ray. Note that image rays can cross each
other which means that one depth point may be in fact be imaged in two positions
on the time migrated section Hubral (1977). Normal rays cross each other if the
curvature of the structure is stronger than the curvature of the wavefront. In that
Chapter 2: Migration: an overview
normal ray
Figure 2.6
13
image ray
The normal ray represents minimum traveltime from a point on the surface to the reflector
and the image ray represents minimum traveltime from a point on the reflector to the
surface. Note that a zero-offset recording has traveled two times (down and up) along
the normal ray path.
case a bow-tie (indicating multiple arrival times) will be observed at the surface.
Due to the smaller curvature of the wavefront at deeper depth levels, a bow-tie is
observed for a smaller curvature of the structure. Note that the imaging path of
the finite difference method is also along the image ray (see apppendix A Hatton
et al., 1990). The discussed finite difference migration and the Kirchhoff integration
method are both methods which perform a temporal wavefield construction, and in
general give a migrated output section in the time domain.
2.4
Migration in terms of deconvolution
A significant improvement in the understanding of the different approaches to migration was introduced by Berkhout and van Wulfften Palthe (1979); Berkhout (1984).
Berkhout (1982) showed that the process of wavefield extrapolation involves a convolution process (forward extrapolation) and a deconvolution process (inverse extrapolation) along the spatial axes. This systems view on migration generated in the
early eighties a fundamentally different view on migration:
➀ The spatial deconvolution process in migration involves a zero-phasing process,
the maximum resolution being given by the bandwidth of the ’spatial wavelet’
in the data. If propagation losses are taken into account, the deconvolution
operator corrects the spatial amplitude spectrum as well (spatial whitening).
➁ The deconvolution process causes diffractor responses to compress, reflection
responses to reposition and reflection amplitudes to adjust. In addition, the
deconvolution process decreases Fresnel zones to their minimum (given by the
frequency content and aperture angle).
➂ For laterally homogeneous media the deconvolution operator is not changing
along one depth level and the deconvolution process can be efficiently applied
by multiplication in the wavenumber domain, yielding the so-called phase shift
method (Gazdag, 1978; Stolt, 1978).
14
2.4 Migration in terms of deconvolution
➃ For laterally inhomogeneous media the deconvolution operator is laterallyvariant and the deconvolution process can be elegantly represented by a matrix
multiplication.
➄ If the Kirchhoff summation approach to migration is applied in a recursive way,
then the discretized, band-limited, recursive deconvolution operator equals the
exact finite-difference operator. In addition, in the wavenumber domain the
recursive Kirchhoff operator equals the phase shift operator.
➅ Since the earth is time-invariant during a seismic experiment, the spatial deconvolution process in seismic migration may always be applied in the temporal
frequency domain, yielding the so-called F-X and F-X-Y algorithms.
➆ The deconvolution formulation to migration yields directly a depth migration
algorithm.
If there are large lateral and vertical velocity changes all time migration approaches
break down. To overcome these problems depth migration need be applied (Berkhout
and van Wulfften Palthe, 1979; Schultz and Sherwood, 1980). The depth migration
method proposed by Berkhout (1982), and based on a spatial convolution and deconvolution process in the space-frequency domain, can handle these lateral velocity
changes in a simple manner. In his depth migration scheme each frequency component of the wavefield is extrapolated to another depth level by means of a spatial
convolution operator. When the velocity changes with lateral position a new convolution (extrapolation) operator is read from a table which is computed in advance
(Blacquière et al., 1989). For an overview of the different implicit and explicit extrapolation operators used in the frequency domain the reader is referred to appendix
A of this thesis.
Another advantage of recursive depth migration is that it automatically handles energy of multiple paths from upper surface points to depth points, while Kirchhoff
depth migration allows a few paths at most to connect a upper surface point with
a depth point. However, the success of depth migration is completely dependent on
the used velocity model, just like any other migration method. Parkes and Hatton
(1987) have shown that positioning errors in the migrated sections are dominated
by inaccuracies in the macro model used in the migration algorithm and not by the
inadequacies of the algorithms themselves. The highest priority must therefore be
assigned to improving methods for estimating the velocity field. The latest developments in depth migration aim at a pre-stack migration technique which can be
combined with a detailed velocity analysis.
At the begin of the 80’s depth migration methods became more popular and due
to the increasing computer power depth migration schemes could also be applied to
pre-stack data (Schultz and Sherwood, 1980). By performing migration directly on
Chapter 2: Migration: an overview
time [s]
-1500
0
-750
lateral position [m]
0
750
1500
0.5
-1500
0
lateral position [m]
0
750
1500
0.5
1.0
1.0
a. shot record at surface
-1500
0
time [s]
-750
15
-750
0
b. extrapolated to 300 m depth
750
0.5
1500
-1500
0
-750
0
750
1500
0.5
1.0
1.0
c. extrapolated to 600 m depth
Figure 2.7
d. extrapolated to 900 m depth
Inverse extrapolation of a diffractor (at 900 [m] depth) response to deeper depth levels
involves a spatial deconvolution step. Note that the effective receiver array becomes
smaller with increasing depths.
the shot records, and not on stacked data, a much better image can be obtained.
The stacking process is based on the assumption of a lateral invariant medium. For
most interesting cases this assumption is not valid and the obtained stack can distort
the quality of dipping reflectors in the depth image. In pre-stack depth migration
the shot records are regarded as the sampling of an upgoing wavefield. Note that
this is only true if all the surface related multiples are removed from the data. This
wavefield is backward propagated in depth by using recursive extrapolation (spatial
deconvolution) schemes. The extrapolation step simulates a downward moving receiver array, as shown in figure 2.7, where a recursive inverse extrapolation algorithm
is used to propagate the wavefield to deeper depth levels.
A very computation intensive but accurate extrapolation method is described by
McMechan (1983); Mufti et al. (1996), where a time reversed finite difference scheme
(Alford et al., 1974) of the two-way wave equation is used to migrate the data. The
process of time reversed migration transforms data from the measurement plane to
the space depth domain by using the seismic data as boundary conditions for the
different time steps. Since the two-way wave equation is used, stable migration of
very steep dips is possible. If the velocities for the migration are chosen correctly,
the wavefield at t = 0, obtained by constructive interference of wavefronts, should
be considered as the final migrated section.
16
2.5
2.5 Inverse scattering
Inverse scattering
Most of the methods described above make use of some type of extrapolation (deconvolution) algorithm to bring the measured data closer to the reflector (scattering
object) and use an additional imaging principle to image the reflector. This extrapolation method is one type of solution of a more general problem of finding the
scattering object from measurements surrounding the object; the so called inverse
scattering problem. The publications to solve for the inverse scattering problem, outside the seismic literature, are extensive. Here only those methods directly related
to the seismic literature are briefly mentioned.
In the inverse scattering theory approach it is explicitly stated that a model is sought,
which is the best in a given sense (e.g. the minimization of some functional defined
over the model space). Tarantola (1984b) uses a linear inverse scattering theory to
solve the seismic problem and has shown that this linearization lead to a solution
strongly related to the Kirchhoff migration method. In the forward modeling, which
is used in the inversion, only the first order scattering energy is taken into account,
multiple reflections are neglected (the so called Born approximation). In another
paper Tarantola (1984a) uses a non-linear approach which strongly resembles the
migration method based on the imaging principle of Claerbout (1971).
2.6
Summary
The desired output of every migration program is a representative image of the
earth. In this image all reflectors are positioned at their correct lateral position
in depth and the interpreter can immediately pinpoint the interesting areas. In
migration however there are several other seismic images involved as shown in figure
2.8 and are given by, the unmigrated shot-record, the stacked section and the time
migrated image (after Parkes and Hatton, 1987). The conventional main product
of processing is obtained by reordering and summing the shot records in the time
domain to a stacked section. Imaging of this stacked section with a time migration
method gives the time migrated section and imaging with a depth migration method
gives an image of the earth. However, the best imaging results are obtained by a
direct map of the (unstacked) shot records to an image of the earth. Unfortunately
pre-stack shot record migration is very computational intensive, specially for 3D
data, to become a standard processing procedure in the near future. Most methods
used today are characterized by a two-stage method; conventional time migration
(finite difference or Kirchhoff) followed by a depth migration along the image rays.
The top right box in figure 2.8 represents a new imaging method based on double
focusing (Berkhout and Rietveld, 1995). The CFP gather represents a half-way
migration result and consists of synthesized shot records. This synthesis process
consist of a weighted stack of the traces in a shot record, where the weights are
Chapter 2: Migration: an overview
synthesis
shot
records
pre-stack depth
migration
nmo and
stack
Figure 2.8
CFP
gather
one-way
image ray
earth
model
normal ray
stacked
section
17
image ray
time
migration
migrated
time
The five seismic domains; the earth is represented by the depth domain, the stacked
seismic section to unmigrated time and the time migrated stack to migrated time. Note
that the domain at the top right is not yet explained, this domain will be introduced in
this thesis.
defined by a solution of the wave equation. The synthesis result defined for one
point in the subsurface is the subject of this thesis and is called Common Focus
Point (CFP) gather. The new method introduces a new image domain called the
one-way time image and has a great affinity with Kirchhoff depth migration as shown
in figure 2.9. However, there is one fundamental difference between the two methods.
In Kirchhoff depth migration the wavefield is focused in one integration step, while
in the CFP method two separate focusing steps are carried out. The result after
one focusing step transforms the shot records to CFP gathers and, as will become
clear in this thesis, this gather is a very suitable domain for velocity (or operator)
analysis.
The extension of the discussed migration techniques to 3-dimensional data is not
straightforward due to the incomplete acquisition at the surface. The Kirchhoff
migration and focusing methods are the most flexible and therefore also the most
often used method at the moment. The depth migration scheme’s require a regular
sampling of the data, which means that a lot of interpolation has to be done (which
increases the already large 3D data-volume even more) before the extrapolation can
be carried out.
To end the historical overview of migration it must be noted that in the first issue of
Geophysics Rieber (1936) describes a method which tries to improve the maximum
sensitivity of the geophone groups. He argues that if a wave arrives at an angle
with respect to the line occupied by the geophone group, it will not reach all of the
geophones at the same time instant, and hence their cumulative impulse will not be
transmitted into the electrical system in phase. The axis of maximum sensitivity of
the geophone group is directed downwards and excludes from the summed record all
18
2.6 Summary
shot records
Kirchhoff
depth
migration
two-way
time image
Figure 2.9
shot records
first focusing
step
second focusing
step
one-way
time image
The difference between Kirchhoff depth migration (left) and CFP migration is in the
division of the focusing step into two processes. The result after the first focusing step
turn out to be a very suitable domain for velocity analysis.
wave components not normal to the geophone group. To improve the sensitivity of
the geophone group Rieber introduced controlled directional sensitivity. In the first
step the individual geophones are measured and in a second step they are combined
not only in their original phase relationship, as is done in ordinary multiple recording, but also in any desired phase relationship. Each trace of the directional analysis
strip represents the sum of the outputs of all the detectors, but with a constant
phase difference introduced between the outputs from adjacent detectors before cumulation. Current processing methods which use Radon transformations (linear,
parabolic, hyperbolic) or inverse ray-tracing are more successful implementations of
the same concepts. The idea of combining weighted receivers at the surface is also
used in the Common Focus Point technology where the weights are chosen such that
the source and receiver sensitivity are focused on one point in the subsurface.
Chapter 2: Migration: an overview
Method
Finite Difference
Kirchhoff
Fourier
F-X
CFP
19
high angles
velocity changes
computational speed
++
++
+
++
++
++
+
+
++
–
+
Table 2.1 A simplified overview of the advantages and disadvantages of the different migration methods discussed in this chapter. In the table ++ means prefered, + well suited, can be
used but not recommended and - means not recommended.
20
2.6 Summary
Chapter 3
Two-way and one-way representations
The seismic analysis tools which are used within the DELPHI group are all based
on the same forward model. This forward model describes the mathematical relationship between the geophysical properties of the earth and the seismic measurements. Once the forward model is defined a related inversion scheme, which estimates the geophysical properties from the seismic measurements, can be derived. In
this chapter the forward model is derived and formulated in such a way that it can
be used in the inversion scheme described in chapter 4.
The representations for both two-way and one-way wavefields are discussed and
compared with each other. Two-way wavefields can be described in terms of the total
acoustic pressure and the total particle velocity. These terms are always coupled by
the two-way wave equations. One-way wavefields are described in terms of waves
traveling in the positive and negative axial direction. If the medium parameters do
not vary in the axial direction the up- and downgoing one-way waves are completely
decoupled; otherwise the coupling between the up- and downgoing waves is expressed
in terms of the axial variations of the medium parameters. Therefore the description
in one-way wavefields is useful when there is a clear prefered direction of propagation.
In surface seismic exploration the vertical direction is regarded as the preferred
direction of propagation, which makes the one-way wave theory very well suited for
seismic applications. At the end of this chapter the forward model used in this thesis
is formulated by making use of the representations of seismic data which are derived
from two- and one-way reciprocity theorems.
3.1
Reciprocity theorems
The aim of seismic wave theory is to solve for the unknown inhomogeneities and the
structural layers in the subsurface of the earth given a measurement at the surface
of the earth. The measured wavefield represents the inhomogeneities in the subsurface of the earth due to the scattering of the incident source wavefield. Solutions
22
3.1 Reciprocity theorems
based on the wave equation try to extract this information, which is limited by
the resolution of the method, from the measured data. However, for many wave
equation-based solutions it is necessary to know the wavefield propagation properties of the subsurface. Unfortunately these propagation properties are part of the
desired information to be extracted from the measured data. Therefore one usually
starts with an initial macro model which describes the global propagation properties of the subsurface. To be able to solve for the unknown, a theorem is needed
which describes the relationship between the measured wavefield and the wavefield
propagation through the macro model. A reciprocity theorem can relate two states
that occur in the same domain in space. These two states can be chosen as the
measured state and the model state, which makes reciprocity theorems fundamental
in seismic wave theory (de Hoop, 1988; Fokkema and van den Berg, 1993). The
reciprocity theorems discussed in this chapter interrelate two acoustic states in a
time-invariant bounded domain D and are used to derive representations for seismic
wavefields. In this section the mathematical tools are introduced and the two-way
and one-way reciprocity theorems are derived. Note that the chapter about symbols
and definitions on page ix describes most of the notation conventions used in this
thesis.
3.1.1
Reciprocity theorem for two-way wavefields
The integral theorem of Gauss interrelates quantities at the surface of a bounded
domain D to quantities inside the domain with the relation
Z
Z
2
∂k Ek d3 x,
(3.1)
Ek nk d x =
∂D
D
where ∂D is the boundary of D, with nk the unit vector normal to ∂D and oriented
away from D, d3 x the elementary volume in 3-dimensional Euclidean space and d2 x
the elementary area of ∂D. Substituting Ek = PA Vk,B − Vk,A PB into equation (3.1)
gives the following integral relation
Z
Z
2
{PA Vk,B − Vk,A PB }nk d x =
∂k {PA Vk,B − Vk,A PB }d3 x.
(3.2)
∂D
D
Equation (3.2) can be used to interrelate two different states, denoted by the subscripts A and B. To use this relation for wavefields the quantities Vk and P occuring
in equation (3.2) must be related to physical parameters which describe the wave
phenomena.
The linearized equation of motion and the linearized equation of deformation are the
coupled equations which are used to describe the propagation of two-way wavefields.
In the frequency domain these equations are given by
∂k P + jωρVk = Fk ,
(3.3)
∂k Vk + jωκP = Q,
(3.4)
Chapter 3: Two-way and one-way representations
23
respectively, where P is the acoustic pressure [Pa = N/m2 ], Vk the particle velocity
[m/s], ρ the volume density of mass [kg/m3 ], κ the compressibility [Pa−1 = m2 /N],
Fk the volume source density of volume force [N/m3 ] and Q the volume source
density of volume injection rate [1/s]. Using the coupled two-way wave equations (3.3) and (3.4) for two different states A and B and the interaction quantity
∂k {PA Vk,B − Vk,A PB } occuring in equation (3.2) gives
Z
{PA Vk,B − Vk,A PB }nk d2 x =
∂D
Z
jω {Vk,A (ρB − ρA )Vk,B − PA (κB − κA )PB }d3 x
ZD
(3.5)
+ {Fk,A Vk,B + QB PA − Fk,B Vk,A − QA PB }d3 x.
D
Equation (3.5) is called Rayleigh’s reciprocity theorem (Rayleigh, 1894; Fokkema and
van den Berg, 1993) of the convolution type, since the products in the frequency domain correspond to convolutions in the time domain. Using the interaction quantity
∗
∂k {PA∗ Vk,B + Vk,A
PB } in equation (3.1) the reciprocity theorem of the correlation
type (Bojarski, 1983) is obtained:
Z
∗
{PA∗ Vk,B + Vk,A
PB }nk d2 x =
Z∂D
∗
−jω {Vk,A
(ρB − ρA )Vk,B + PA∗ (κB − κA )PB }d3 x
D
Z
∗
∗
Vk,B + QB PA∗ + Fk,B Vk,A
+ Q∗A PB }d3 x.
(3.6)
+ {Fk,A
D
These scalar-wave reciprocity theorems are used to derive wavefield representations
of seismic data and the related forward model. Note that by eliminating Vk from
equation (3.3) and equation (3.4) the wave equation in the frequency domain is
obtained
1
(3.7)
ρ∂k ( ∂k P ) + ω 2 c−2 P = −s
ρ
1
with the acoustic velocity c = (κρ)− 2 [m/s] and the source term s = jωρQ −
ρ∂k ( ρ1 Fk ).
3.1.2
Reciprocity theorem for one-way wavefields
In surface seismics the prefered direction of propagation is along the x3 (vertical)
axis; it is therefore useful to reorganize equation (3.3) and equation (3.4) in such a
way that the ∂3 derivatives are separated from the ∂1 , ∂2 derivatives. Eliminating
V1 and V2 from equations (3.3) and (3.4) gives the desired result
∂3 Q = AQ + D,
(3.8)
24
3.1 Reciprocity theorems
with the two-way wave vector
P
,
V3
(3.9)
!
F3
,
1
∂α ( ρ1 Fα )
Q − jω
(3.10)
Q=
the source vector
D=
and the two-way matrix operator
A=
0
1
jωρ H2
!
−jωρ
.
0
(3.11)
In the two-way matrix operator of equation (3.11) H2 represents the Helmholtz
operator which is given by
H2 =
ω 2
c
1
+ ρ∂α ( ∂α ·).
ρ
(3.12)
The decomposition of the acoustic two-way wave equation (3.8) into equations for upand downgoing one-way wavefields is carried out by the inverse of matrix operator
L. The matrix operator L is used in the diagonalization of the matrix operator A
such that Λ in
A = −jωLΛL−1
(3.13)
is a diagonal operator matrix (see Fishman et al. (1987); for references in the seismic
context see Wapenaar and Berkhout (1989); de Hoop (1992)). The composition from
the one-way wave vector P , which contains downgoing P + and upgoing P − waves,
to the two-way wave vector Q is then defined by
P
Q=
= LP ,
(3.14)
V3
+
P
where P =
. In a similar way the source vector D is composed from up- S −
P−
and downgoing S + sources according to
D = LS,
(3.15)
+
S
. Note that the composition matrix L can be chosen flux norS−
malized (Ursin, 1983; Wapenaar, 1996c) which is convenient in the derivation of
reciprocal relations. Substituting the composition equations (3.14) and (3.15) into
equation (3.8) gives the coupled one-way wave equations
where S =
∂3 P = BP + S,
(3.16)
Chapter 3: Two-way and one-way representations
25
where the one-way operator matrix B is given by
B = −jωΛ + Θ,
(3.17)
∗+
0
(3.18)
with the diagonal matrix Λ by
Λ=
0
,
∗−
where ∗+ = −Λ− . The coupling operator matrix Θ is defined as
+
T
R−
−1
Θ = −L ∂3 L =
,
−R+ −T −
(3.19)
where R± and T ± are reflection and transmission operators. From the structure
of equations (3.16) to (3.19) it follows that −jωΛ accounts mainly for (downward/upward) propagation and Θ for scattering mainly due to vertical variations of
the medium parameters. This explicit distinction between propagation and scattering is an important advantage of the one-way wave equation over the two-way wave
equation.
For a homogeneous medium the coupling operator matrix Θ = O and the diagonal
matrix Λ can be written in the wavenumber frequency domain as
−jkz
0
,
(3.20)
−jω Λ̃ =
0
jkz
where the vertical wavenumber
kz = ω˜
∗+ = −ω˜∗− =
(√
k 2 − kα kα
√
−j kα kα − k 2
kα kα ≤ k 2
kα kα > k 2
.
(3.21)
A solution of the one-way wave equation (3.16) in a homogeneous medium is given by
the well known phase shift operator (Gazdag, 1978). For a more detailed discussion
about the properties of the one-way wave equation the reader is referred to Fishman
et al. (1987); Wapenaar and Berkhout (1989); Wapenaar and Grimbergen (1996).
Starting from the one-way wave equation (3.16) and the interaction quantity
∂3 {PA+ PB− − PA− PB+ }
(3.22)
the reciprocity theorem of the convolution type for one-way wavefields is derived.
Note that the interaction quantity (3.22) contains waves that propagate in opposite
directions. To simplify the notation this interaction quantity is rewritten to
+ 0 1
PB
+ −
= ∂3 {P TA NP B }.
(3.23)
∂3 PA PA
−1 0
PB−
26
3.1 Reciprocity theorems
∂D
x1
x2
D
x3
∂D
Figure 3.1
n = (0, 0, −1)
n = (0, 0, 1)
Configuration for one-way reciprocity theorems. D is a volume enclosed by two infinite
parallel surfaces normal to the x3 axis with the outward pointing vector n defined on
∂D. The combination of these two surfaces is denoted by ∂D.
In figure 3.1 the domain used in the one-way reciprocity theorem is shown. The
domain D is a volume ’enclosed’ by two infinite parallel surfaces normal to the x3
axis. The combination of these two surfaces is denoted by ∂D. Substituting equation
(3.23) into equation (3.2) with the domain configuration of figure 3.1 and by making
use of a specific symmetry property of the one-way operator matrix B (Wapenaar,
1996b; Wapenaar and Grimbergen, 1996) yields the following one-way reciprocity
theorem of the convolution type
Z
P TA NP B n3 d2 x =
∂D
Z
P TA N(BB − BA )P B d3 x
ZD
+ {P TA NS B + S TA NP B }d3 x.
(3.24)
D
The clear distinction between propagation and scattering in the one-way wave equation can be observed by substituting equation (3.17) into equation (3.24) yielding
Z
P TA NP B n3 d2 x =
∂D
Z
1 T
P A N(ΘB − ΘA )P B }d3 x
−jω {P TA N(ΛB − ΛA )P B −
jω
D
Z
+ {P TA NS B + S TA NP B }d3 x.
(3.25)
D
Note that the one-way reciprocity theorem of the convolution type has the same
form as the two-way reciprocity relation given in equation (3.5); a surface integral
over the interaction quantity on one side and volume integrals containing a contrast
function and sources on the other side. The contrast term (BB − BA ) has the
Chapter 3: Two-way and one-way representations
27
property that it vanishes when the medium parameters in state A and state B are
identical. In equation (3.25) it can be seen that the contrast functions allow an
independent choice of Λ (propagation) and Θ (scattering).
The one-way reciprocity theorem of the correlation type is obtained by using the
interaction quantity
+ 0
PB
+
− ∗ 1
∂3 PA PA
= ∂3 {P H
(3.26)
A JP B }
0 −1
PB−
in Gauss’s theorem which results in
Z
2
PH
A JP B n3 d x ≈
∂D
Z
3
PH
A J(BB − BA )P B d x
D
Z
H
3
+ {P H
A JS B + S A JP B }d x.
(3.27)
D
The approximation sign in equation (3.27) refers to the use of an approximation
in the derivation of a modified symmetry property of the one-way operator matrix
B. In this approximation evanescent wave modes are erroneously handled. For
a detailed comparison between the one-way and two-way reciprocity theorem and
the symmetry properties of the one-way operator matrix B the reader is referred
to Wapenaar (1996b). The one-way reciprocity theorems of the convolution and
correlation type are used to derive one-way representations of seismic data.
3.2
Integral representations for two-way wavefields
An integral representation expresses a wavefield quantity at some point in a medium
in terms of boundary and volume integrals over the wavefield, the source distribution and the contrast function. Integral representations for two-way wavefields are
derived by using the reciprocity theorems derived in section 3.1.1 and taking some
special choices for the two states A and B in the reciprocity relation. Acoustic
representations have been introduced by Lord Rayleigh (Rayleigh, 1894). For the
derivation of a general representation of an acoustic wavefield the domain D consist
of a reference medium {κ̄(x), ρ̄(x)} where a contrasting medium Ω {κ(x), ρ(x)} is
embedded. The wavefield of state A in the reciprocity theorem is the Green’s wavefield propagating in the reference medium with a point source at x′ inside D. Note
that the wavefield described by the Green’s function is chosen to be causally related
to the source function. For state B the wavefield in the actual medium is chosen,
with a source position which lies outside the scattering part Ω of domain D and the
observation point at x is a point inside D.
Substituting these states, which are summarized in table (3.1), into the reciprocity
28
3.2 Integral representations for two-way wavefields
State A
∂k G(x, x′ ) + jω ρ̄(x)Γk (x, x′ ) = 0
∂k Γk (x, x′ ) + jωκ̄(x)G(x, x′ ) = δ(x − x′ )
State B
∂k P (x) + jωρ(x)Vk (x) = 0
∂k Vk (x) + jωκ(x)P (x) = S(x)
Table 3.1 States for a general representation of two-way wavefields.
theorem of the convolution type (equation (3.5)) gives the following general representation
I
{G(x, x′ )Vk (x) − Γk (x, x′ )P (x)}nk d2 x =
∂D
Z
jω {Γk (x, x′ )(ρ(x) − ρ̄(x))Vk (x) − G(x, x′ )(κ(x) − κ̄(x))P (x)}d3 x
ZD
S(x)G(x, x′ )d3 x − P (x′ ).
(3.28)
+
D
By using G(x, x′ ) = G(x′ , x), ∆ρ(x) = ρ(x) − ρ̄(x), ∆κ(x) = κ(x) − κ̄(x), ∂k nk =
−1
′
∂n , Γk (x, x′ ) = jω−1
ρ̄(x) ∂k G(x, x ), Vk (x) = jωρ(x) ∂k P (x) and interchanging x with
x′ gives
P (x) =
Z
G(x, x′ )S(x′ )d3 x′ +
D
I
1
1
1
G(x, x′ )∂n′ P (x′ ) −
(∂ ′ G(x, x′ )) P (x′ )}d2 x′ +
{
jω ∂D ρ(x′ )
ρ̄(x′ ) n
Z
∆ρ(x′ )
1
(∂ ′ G(x, x′ )) ∂k′ P (x′ ) + ω 2 ∆κ(x′ )G(x, x′ )P (x′ )}d3 x′ .
{
jω D ρ̄(x′ )ρ(x′ ) k
(3.29)
This integral representation is known in the acoustic literature as the KirchhoffHelmholtz integral (see, for instance Schneider, 1978; Clayton and Stolt, 1981;
Berkhout, 1987). The three terms on the right hand side of equation (3.29) can
be interpreted as follows (see also Figure 3.2): The first term ➀, a volume integral
over the source domain, represents the direct field traveling from the source distribution to the observation point at x. The second term ➁, a surface integral over
∂D, represents the contributions from sources and scatterers outside the domain
D. The third term ➂, a volume integral over the domain D, can be interpreted as
the contribution from scattering objects inside D. If the actual medium ρ(x), κ(x)
differs from the reference medium ρ̄(x), κ̄(x) this volume integral has a contribution
to the field P (x). The strength of this contribution is dependent on the contrast
between the actual medium and the reference medium.
Define the incident field P i as the wavefield in the reference medium and the
scattered wavefield P s as the difference of the total wavefield P and the incident
Chapter 3: Two-way and one-way representations
P (x)
➀
➁
➁
D
κ̄ ρ̄
Figure 3.2
S(x′ )
➀
➂
∂D
29
Ω
κρ
General representation of the wavefield in domain D with a scattering area Ω, where x
represents the point of observation and x′ represents the source position inside D. Note
that is is possible to distinguish three different contributions at the point of observation.
field according to
P s (x) =P (x) − P i (x), with
Z
G(x, x′ )S(x′ )d3 x′ .
P i (x) =
(3.30)
D
An integral representation of the scattered wavefield can be obtained by assuming
that the surface ∂D in equation (3.29) is a sphere with infinite radius, so that
the surface integral ➁ vanishes on account of Sommerfeld’s radiation condition,
(Bleistein, 1984)). The integral representation for the scattered wavefield is then
given by
Z
∆ρ(x′ )
1
(∂ ′ G(x, x′ )) ∂k′ P (x′ ) + ω 2 ∆κ(x′ )G(x, x′ )P (x′ )}d3 x′ .
{
P s (x) =
jω Ω ρ̄(x′ )ρ(x′ ) k
(3.31)
According to this representation the scattering object can be replaced by a set of
distributed sources, which are the product of the contrasting terms and the total
field inside the scatterer. This representation of the scattered field will be used in
the derivation of the double focusing procedure described in chapter 4.
Another useful representation for the wavefield inside D can be obtained by assuming that the domain D has a finite extent and does not contain any contrasts, so
the volume integral with the contrast terms ∆κ and ∆ρ in equation (3.29) vanish
completely. The resulting representation of the scattered field is then given by
I
1
1
P s (x) =
{G(x, x′ )∂n′ P s (x′ ) − (∂n′ G(x, x′ )) P s (x′ )}d2 x′ .
(3.32)
jω ∂D ρ(x′ )
30
3.2 Integral representations for two-way wavefields
x′ ∂D
∂D
∗
G (x, x′ )
x
G(x, x′ )
a.
Figure 3.3
x
b.
x′
a) Forward propagating version of the Kirchhoff integral; the Green’s function propagates inward starting from the surface ∂D. b) Backward propagating version of the Kirchhoff integral; the Green’s function propagates outward starting from the surface ∂D.
In equation (3.32) x′ can be interpreted as the position of a secondary source located
on the surface ∂D. This secondary source has a strength equal to the wavefield at x′
and is determined by the wavefields originating from sources and scattering objects
located outside D. The integral representation describes forward propagation of the
measured wavefield at the surface towards a position x inside the medium D. If
parts of the surface ∂D are extended to infinity the contribution of these surfaces
vanishes on account of Sommerfeld’s radiation condition.
By replacing G with the complex conjugate of G∗ in equation (3.32) the backward
propagating version of the Kirchhoff integral is obtained (Bojarski, 1983; Wapenaar
et al., 1989). Note that G∗ is also a solution of the two-way wave equation and
represents the anti-causal Green’s wavefield. Replacing G with G∗ in equation (3.32)
gives
P s (x) =
1
jω
I
∂D
1
{G∗ (x, x′ )∂n′ P s (x′ ) − (∂n′ G∗ (x, x′ )) P s (x′ )} d2 x′ .
ρ(x′ )
(3.33)
The integral representation describes backward propagation of the measured field at
the surface towards a position x inside the medium D. If parts of the surface ∂D
are extended to infinity the contribution of these surfaces will not vanish. The Sommerfeld’s radiation condition requires that the Green’s function propagates outward
through ∂D, in the same direction as the total wavefield and this condition is not
satisfied for the backward propagation Green’s function.
Equations (3.32) and (3.33) are the basis for forward and inverse wavefield extrapolation techniques and are graphically illustrated in figure 3.3. By choosing appropriate (fully reflecting) boundary conditions for the Green’s function in equation (3.32)
and (3.33) the Rayleigh I (rigid boundary) and Rayleigh II (compliant boundary)
integrals for forward and inverse extrapolation can be derived (see Berkhout, 1987;
Berkhout and Wapenaar, 1989).
Chapter 3: Two-way and one-way representations
3.3
31
Integral representations for one-way wavefields
Integral representations for one-way wavefields are derived by using the reciprocity
theorem derived in section (3.1.2) and taking some special choices for the two states
A and B in the reciprocity relation. The domain D is a volume ’enclosed’ by two
infinite parallel surfaces normal to the x3 axis, see figure 3.1.
State A
¯
∂3 G(x, x′ ) − B(x)G(x,
x′ ) = Iδ(x − x′ )
State B
∂3 P (x) − B(x)P (x) = S(x)
Table 3.2 States for general representation of one-way wavefields.
For the derivation of a general one-way representations state A in the reciprocity
theorem is replaced by a one-way Green’s function and state B by the actual one-way
wavefield. The Green’s function is defined with reference operator B̄ and a point
source at x′ inside D. In the acoustic one-way wave equation the Green’s operator
matrix defined by these choices is given by
′
G(x, x ) =
G+,+
G−,+
G+,−
(x, x′ ),
G−,−
(3.34)
where the superscripts refer to the propagation direction at x and x′ respectively, see
also figure 3.4. For state B, with the actual operator B, the source is positioned inside
D at S(x). These two states, which are used to derive the general representation,
are summarized in table (3.2).
Substituting these two states into the reciprocity theorem of the convolution type,
G+,+
G−,+
Figure 3.4
G+,−
G−,−
The elements in the one-way Green’s function. Note that the Green’s functions for the
direction converted waves G+,− and G−,+ are only non-zero if Θ̄(x) 6= 0.
32
3.3 Integral representations for one-way wavefields
given by equation (3.25), gives the following expression;
Z
GT (x, x′ )NP (x)n3 d2 x =
∂D
Z
GT (x, x′ )N(Λ(x) − Λ̄(x))P (x)d3 x
−jω
ZD
GT (x, x′ )N(Θ(x) − Θ̄(x))P (x)d3 x
+
D
Z
+
GT (x, x′ )NS(x)d3 x + NP (x′ ).
(3.35)
D
This equation can be simplified by making use of a reciprocity relation between two
one-way Green’s functions (Wapenaar, 1996a). This relation is given by
GT (x, x′ ) = −NG(x′ , x)N−1
(3.36)
and is obtained by applying the reciprocity relation of the convolution type (3.27)
to two Green’s states defined in the same medium, where the wavefield in state A is
given by G(x, x′ ) and the wavefield of state B by G(x, x′′ ). The medium outside
D is chosen homogeneous and isotropic. Note that this reciprocity relation for the
Green’s function holds due to the flux normalization of the composition operator
matrices L.
Multiplying equation (3.35) with N−1 , using the reciprocity relation of equation
(3.36) and interchanging x and x′ gives the following representation for the one
wavefield at x
Z
P (x) =
G(x, x′ )S(x′ )d3 x′
ZD
−
G(x, x′ )P (x′ )n3 d2 x′
∂D
Z
G(x, x′ )(Λ(x′ ) − Λ̄(x′ ))P (x′ )d3 x′
−jω
D
Z
G(x, x′ )(Θ(x′ ) − Θ̄(x′ ))P (x′ )d3 x′ .
(3.37)
+
D
Note the analogy with the two-way representation of equation (3.29). The right
hand-side of equation (3.37) contains respectively, a direct wave contribution, a
boundary integral over the interaction quantity, a volume integral over the contrast
term for propagation and a volume integral over the contrast term for scattering.
This general one-way representation is used to derive volume and surface integral
representations of the scattered one-way wavefield. A volume integral representation
is obtained by letting the boundary integral over ∂D vanish by choosing domain D
equal to R3 . Further subtracting the incident wavefield
Z
P i (x) =
G(x, x′ )S(x′ )d3 x′
(3.38)
D
Chapter 3: Two-way and one-way representations
from the result of equation (3.37) gives
Z
G(x, x′ )(Λ(x′ ) − Λ̄(x′ ))P (x′ )d3 x′
P s (x) = −jω
ZD
G(x, x′ )(Θ(x′ ) − Θ̄(x′ ))P (x′ )d3 x′ ,
+
D
Z
=
G(x, x′ )(B(x′ ) − B̄(x′ ))P (x′ )d3 x′ ,
33
(3.39)
D
which represents the scattered one-way wavefield due to a contrasting medium in
domain D. The operators Θ̄ and Λ̄ for the Green’s state can be chosen independently
from each other, for example by choosing Θ̄ = O and Λ̄ = Λ only one integral
remains in equation (3.39).
Another useful representation is obtained when the domain D is source free and does
not contain any contrasts. Then all the volume integrals in equation (3.37) vanish
which leaves
Z
s
G(x, x′ )P s (x′ )n3 d2 x′ .
(3.40)
P (x) = −
∂D
Note that the most important contribution at x is the integral over that part of the
surface ∂D which is positioned between x and the source position xs (see figure 3.5).
If the upper half-space, shown in figure 3.5, is chosen source free (outside D) and homogeneous and xs is chosen below ∂D0 then the contribution of the surface integral
over ∂D1 vanishes and only the surface integral over ∂D0 remains. This resulting
equation (3.40) describes ’one-way wavefield extrapolation’, as it is commonly named
in the literature on seismic exploration (see for example Claerbout, 1976; Berkhout,
1982). Another way to derive equation (3.40) is to substitute P = P + + P − into
the two-way representation of equation (3.32) and rewriting the resulting expression
(see for example Berkhout and Wapenaar, 1989).
The backward propagating version of equation (3.40) can be obtained by using the
reciprocity relation of the correlation type (equation (3.27)) and following the same
steps as before which results in
Z
KG∗ (x, x′ )K(Λ(x′ ) − Λ̄(x′ ))P (x′ )d3 x′
P s (x) ≈ −jω
ZD
+
KG∗ (x, x′ )K(Θ(x′ ) − Θ̄(x′ ))P (x′ )d3 x′ ,
D
Z
=
KG∗ (x, x′ )K(B(x′ ) − B̄(x′ ))P (x′ )d3 x′ ,
(3.41)
D
representing the backpropagating counterpart of equation (3.39) with
0 1
−1
−1
K=J N=N J=
.
1 0
(3.42)
34
3.3 Integral representations for one-way wavefields
P − [G+,+ ]∗
P − G−,−
∂D1
∂D1
x
P − G−,−
∂D0
a.
Figure 3.5
x
P − [G+,+ ]∗
∂D0
xs
b.
xs
Forward (a) and inverse (b) extrapolation with the forward and backward propagating
Green’s function. Note that for forward propagation the surface integral over ∂D1 vanishes and the only contributions comes from the surface integral over ∂D0 . For inverse
extrapolation the surface integral over ∂D0 does not vanish and represents the evanescent waves. The propagating waves are represented by the surface integral over ∂D1 .
Using now the same domain as used in the derivation of equation (3.40) yields
Z
KG∗ (x, x′ )KP s (x′ )n3 d2 x′ ,
(3.43)
P s (x) ≈ −
∂D
In equation (3.43) the operators on the diagonal of G∗ can be interpreted as an
approximation of the inverse operators of the diagonal elements in the Green’s function G, since forward and backward propagation of the one-way wavefield over the
same distance should yield the original input wavefield. The exact inverse of the
Green’s operator cannot be used in numerical implementations because it blows up
the evanescent fields. In the ’matched filter’ approximation (equation (3.43)) the
evanescent modes are erroneously handled. Although this limits the maximum obtainable spatial resolution, it also assures stability since the evanescent wave modes
are suppressed instead of amplified, Wapenaar and Berkhout (1989). In the remainder of this thesis the approximation sign ≈ is replaced with an equal = sign if
only the evanescent waves are neglected.
The most important contribution in the backward propagating solution at x is the
surface integral over ∂D1 if the source is positioned below ∂D0 . If it would be
possible to use the exact inverse of the Green’s operator then the contribution of the
surface ∂D1 would be sufficient to represent the field inside D. However by using
the matched filter approach the surface integral over ∂D0 must be included to get
a correct representation of the evanescent wavefield. Usually one is only interested
in the propagating wavefield and the surface integral over ∂D0 can be neglected. In
figure 3.5 the lines starting at the boundaries indicate the propagation direction of
the Green’s wavefields G−,− and [G+,+ ]∗ . The gray lines refer to the contribution
of the evanescent waves and the solid lines refer to the propagating wavefield. The
actual wavefield P − (x) originates from a source below the surface ∂D0 . Note that
if one is only interested in the propagating wavefield the contribution from the
surface ∂D0 can be neglected in the backward propagating integral representation. A
Chapter 3: Two-way and one-way representations
35
more detailed discussion of the approximations involved in the backward propagating
version of the two-way and one-way representation can be found in Wapenaar and
Berkhout (1989).
Finally all the tools are ready to derive the forward model used in this thesis. The
forward model for (primary) scattered data is derived by using the states indicated
in table (3.3) in the representation of equation (3.37).
State A
∂3 G(x, x′ ) + jωΛ(x)G(x, x′ ) = Iδ(x − x′ )
State B
∂3 P (x) − B(x)P (x) = S(x)δ(x − xs )
Table 3.3 Reciprocity states for scattered one-way wavefields.
The propagation operator Λ̄ = Λ for the Green’s state is chosen the same as in the
actual medium and the scattering operator for the Green’s state is chosen equal to
zero, Θ = O. The Green’s function used in the representation is causally related to
the source function. With these choices the Green’s matrix G(x, x′ ) contains only
diagonal elements which are given by
H(x3 − x′3 )Wp+ (x, x′ )
0
′
,
(3.44)
G(x, x ) =
0
−H(x′3 − x3 )Wp− (x, x′ )
where H(x3 ) is the Heaviside step function. Wp+ (x, x′ ) and Wp− (x, x′ ) are referred
to as the primary propagators in the actual medium for downgoing and upgoing
waves. Note that for a homogeneous medium the primary propagators can be written in a closed form. These closed form expressions are useful in the numerical
implementation of the forward model which is sketched in the next section. For nonhomogeneous media the primary propagator can be built up recursively (Berkhout,
1982; Wapenaar, 1996a).
The actual wavefield P in the representation originates from the downgoing source
+
S0
S(x) =
(x).
(3.45)
0
The domain D in equation (3.37) spans the entire space R3 in which the upper halfspace x3 ≤ x3,m is homogeneous and the lower half-space x3 > x3,m is inhomogeneous. The source and receiver are positioned in the homogeneous upper half-space.
In this configuration the surface integral in equation (3.37) vanishes which leaves
the following expression for a detector position at xr in the homogeneous halfspace
Z
G(xr , x)Θ(x)P (x)d3 x,
(3.46)
P (xr ) = P i (xr ) +
Ω
i
where P (xr ) represents the incident field defined by equation (3.38). Writing equation (3.46) into up- and downgoing components by using the Green’s matrix of
36
3.3 Integral representations for one-way wavefields
homogeneous half-space
S0+ (xs )
P − (xr )
x1
x2
x3
x3,0
Wp+ (x′ , xs )
Wp− (xr , x)
x3
R+ (x, x′ )
Ω
Figure 3.6
Configuration for scattered one-way wavefields using primary propagators Wp+ (x′ , xs )
and Wp− (xr , x) in the medium between the source and receiver positions and the contrast medium.
equation (3.44) gives for the upgoing wave at xr
P − (xr ) − P −,i (xr ) =
Z
ZΩ
Ω
Wp− (xr , x)R+ (x)P + (x)d3 x+
Wp− (xr , x)T − (x)P − (x)d3 x,
(3.47)
where Ω denotes the lower half-space x3 > x3,0 , see figure 3.6. The upgoing scattered
wavefield P −,s = P − − P −,i at xr consists of two elements; the transmitted upgoing wavefield T − P − and the reflected downgoing wavefield R+ P + originating from
the inhomogeneities in the domain Ω. Note that the reflection operator R+ (x) in
equation (3.47) can be represented by the integral
R+ (x)P + (x) =
Z
IR2
+
R (x, x′ )P + (x′ ) x′ =x3 d2 x′ ,
(3.48)
3
where R+ (x, x′ ) is the kernel of operator R+ (x), with x′ = (x′1 , x′2 , x′3 = x3 ) a
point defined on the horizontal reflector, see figure 3.6. An expression for primary
scattered data is obtained by replacing P (x) in the right-hand side of equation
(3.46) with P +,i (x) = Wp+ (x, xs )S0+ (xs ) and P − (x) = 0. In this replacement the
field P (x) which is represented completely by equation (3.37) is approximated by
the incident wavefield (the first right hand-side term in equation (3.37)). Using this
Chapter 3: Two-way and one-way representations
37
approximation in equation (3.47) gives for the upgoing scattered wavefield
Z Z
P −,s (xr , xs ) =
Wp− (xr , x) R+ (x, x′ )Wp+ (x′ , xs ) x′ =x S0+ (xs )d2 x′ d3 x.
Ω
IR2
3
3
(3.49)
downsource
wavefield from the source position downwards to the scattering medium at x′ , here
the wavefield gets reflected by R+ (x, x′ ) which is represented by the surface integral
over ∂Ω, finally the reflected wavefield is propagated upwards by the propagation
operator Wp− to a receiver position at xr . This useful representation of seismic
data was first introduced in a discrete formulation by Berkhout (1982) for acoustic
one-way wavefields.
Equation (3.49) can be read from right to left as follows; S0+ represents the
going source function at xs , the propagation operator Wp+ propagates the
An alternative more mathematical interpretation of equation (3.49) is that it represents the first order term of a Bremmer series (Corones, 1975; Wapenaar, 1996a).
The Bremmer series is a straightforward one-way representation of primary and multiple reflection data. In the Bremmer series every term represents an independent
part of the wavefield, while in a Neumann series (which is used in two-way methods) the wavefield represented by the first term is also adjusted by the higher order
terms. Note that the one-way reflection operator R+ is proportional to the vertical variations of the medium parameters which makes it very suitable for seismic
applications because in seismic applications the vertical variations are much more
pronounced than the horizontal variations. Compared with the two-way representation, where a contrast term is defined between the actual and the background
medium, the one-way representation is more convenient for layered media. Note
also that it is possible to replace in representation (3.49) the primary propagators
Wp± (x, x′ ) by generalized primary propagators in which internal multiple scattering
is taken into account in a consistent manner (see Wapenaar, 1996a). So the representation of equation (3.49) is not limited to simple configurations, complicated
propagation effects (for example due to fine layering) can be taken into account by
the W operators and reflection is included by the R operator.
3.4
The WRW model in matrix notation
Although integral representations of seismic data are sometimes more appealing, in
real life seismic measurements are always sampled on discrete ’points’ in space and
time and therefore a matrix notation is more convenient. Another advantage of
the matrix notation is that it makes the numerical implementation of the derived
scheme straightforward. The matrix notation used in this thesis is based on the
notation of Berkhout, (Berkhout, 1982, chapter VI). For a detailed discussion of the
matrix representation of seismic data and the matrix notation the reader is referred
to Appendix B.
38
3.4 The WRW model in matrix notation
S0+ (zs )
P − (zr , zs )
W+ (zm , zs )
W− (zr , zm )
R+ (zm )
Figure 3.7
WRW scheme for one seismic experiment and one reflector.
The representation of scattered one-way wavefields given by equation (3.49) is written in matrix notation as
X
P − (zr , zs ) =
W− (zr , zm )R+ (zm )W+ (zm , zs )S0+ (zs ),
(3.50)
m
where the elements in the vector P − (zr , zs ) represent the discrete detector positions
of one seismic experiment for one frequency. The propagation matrices W± and the
reflection matrix R+ are discrete version of the kernels in equation (3.49). Figure 3.7
shows a block diagram of equation (3.50). The extension of the single experiment
expression of equation (3.50) to a multiple experiment expression is easily done by
combining the individual experiments P − (zr , zs ) into one matrix P− (zr , zs ) in which
every column represents one experiment (see also Appendix B).
The propagation matrices represented by W± can be implemented in several ways.
One of the implementations makes use of a recursive extrapolation scheme in the
space-frequency domain. In this recursive extrapolation scheme a spatial convolution
operator is used which extrapolates the data from one depth level to another depth
level (where the depth step is in general small compared with the wavelength). This
recursive extrapolation scheme can be represented in matrix notation by
W+ (zm , zs ) = W+ (zm , zm−1 ) . . . W+ (z1 , z0 )W+ (z0 , zs ),
(3.51)
where every matrix W+ represents an extrapolation step from one depth level to
another. The columns of the W+ matrices contain the spatial convolution operators.
These operators are calculated only once for the velocity and frequency range of
interest and stored in a table. During the extrapolation the operator needed for the
position of interest is read from the table (Blacquière, 1989).
Note that within the spatial operator length the medium is assumed to be homogeneous, with the velocity given by the central point of the operator. If the medium
varies strongly in the lateral direction this assumption is pushed to its limits (Grimbergen et al., 1996). Therefore in the design of the spatial convolution operators the
Chapter 3: Two-way and one-way representations
D− (zr )
P(zr , zs )
R− (zs , zr )
+
39
D+ (zs )
S0 (zs )
W+ (zm , zs )
W− (zr , zm )
R+ (zm )
Figure 3.8
WRW forward modeling scheme for seismic data including the surface related multiples.
aim is to make the length of the operators as short as possible for a certain maximum
angle of propagation. For a detailed discussion about the calculation of these short
optimized extrapolation operators for 2- and 3-dimensional extrapolation the reader
is referred to appendix A.
Starting from the scheme given in figure 3.7 it is very easy to add the influence
of surface-related multiples. The multiple free response X0 (zr , zs ) is defined as the
total spatial impulse response of the lower half space z > z0 without the multiples
related to depth level z0 . So with this definition equation (3.50) can be rewritten as
P − (zr , zs ) = X0 (zr , zs )S0+ (zs ),
(3.52)
The surface-related multiples are represented by a feed back loop from the left
branch of the scheme to the right branch (Verschuur, 1991) scaled by the reflection
operator R− (z0 ) at the surface of the earth. The total spatial impulse response of
the lower half space z > z0 that includes all the multiples related to depth level z0
is written as
−1
X(zr , zs ) = I − X0 (zr , zs )W+ (zs , z0 )R− (z0 )W− (z0 , zr )
X0 (zr , zs ),
or
−1
X(zr , zs ) = I − X0 (zr , zs )R− (zs , zr )
X0 (zr , zs ),
(3.53)
Equation (3.53) forms the basis of the surface-related multiple removal scheme developed by Berkhout (1982) and Verschuur (1991). Note that in this equation the
generation of surface-related multiples is explicitly described. The relation between
the reflection matrix R− (z0 ) and R− (zr , zs ) is that the latter includes the propagation from the surface z0 to the source and receiver positions.
The decomposition of the source and the received wavefield into up- and downgoing
waves and the directivity of the source and receiver arrays can be included by adding
the operators D± in the scheme. These surface-related operators and the feed back
loop for the surface-related multiples are included in the block diagram of figure 3.8.
40
P(zr , zs )
3.4 The WRW model in matrix notation
D− (zr )
R− (zr , zs )
+
D+ (zs )
S0 (zs )
X0 (zr , zs )
X(zr , zs )
Figure 3.9
Full WRW forward modeling scheme for seismic data, which includes all surface-related
and internal multiples.
The relation between the multiple free data and the data which includes all the
surface-related multiples is shown in its most simple form in figure 3.9 and is represented by
P(zr , zs ) = D− (zr )X(zr , zs )D+ (zs )S0 (zs ),
(3.54)
where the vector P is replaced with matrix P to represent the multiple shot experiment. .
Figure 3.9 shows the complete forward modeling scheme of seismic data. The same
scheme can also be used for the representation of elastic wave propagation (Wapenaar and Berkhout, 1989). For elastic wave propagation the D operators are (de)
composition operators which decompose the elastic wavefield into up- and downgoing potentials for compressional and shear wavefields.
Chapter 4
Imaging by double focusing
The main goal in seismic data processing is to image the geological structures in
the subsurface of the earth. In this chapter it will be shown that the procedure of
imaging seismic data can be split into two focusing steps. The intermediate result
which is obtained after one focusing step is in a very suitable domain for velocity and
AVO analysis. The second focusing step yields a representation of the seismic image,
which can be interpreted by a geologist to identify the geological structures of the
subsurface. In this chapter the double focusing procedure is explained for two-way
and one-way wave fields. However, the one-way solution is prefered because it is
more convenient for seismic wave fields.
The two- and one-way representations derived in the previous chapter are used
in this chapter to explain the double focusing procedure. In the first section the
representations are used to describe the two-way inverse scattering problem. It will
be shown that the inverse scattering problem is represented by a non-linear equation.
In the last sections a solution of this non-linear equation is expressed in a double
focusing procedure for two-way and one-way wave fields. In the double focusing
procedure the backward propagating version of the Kirchhoff integral of equation
(3.33) plays an important role.
4.1
Inverse scattering problem
The integral equation formulation of the scattering problem can be written into two
related equations which are based on equations (3.31) and (3.30). In order to keep
the notation simple the density is taken to be constant everywhere in space (ρ = ρ̄),
so the density contrast in equation (3.31) vanishes. The medium configuration which
is used to describe the scattering problem is shown in figure 4.1, where the scattering
domain is denoted by Ω. The source positioned at xs is located inside D, but outside
Ω. The Green’s function G(x, x′ ) satisfies the Helmholtz equation defined in a
background medium with appropriate boundary conditions. The resulting equation
42
4.1 Inverse scattering problem
xr
xs
D
∂D
κ ρ̄
κ̄ ρ̄
Figure 4.1
Ω
General representation of the wave field in domain D with scattering domain Ω.
represents the measured data at a point xr outside the scattering domain Ω and is
given by
χ(xr )P (xr , xs ) = P i (xr , xs ) + P s (xr , xs ),
(4.1)
where P (xr , xs ) represents the measured data at the receiver positions xr due to a
source at xs . In equation (4.1) P i (xr , xs ) represents the incident field and P s (xr , xs )
the scattered field which is given by
Z
P s (xr , xs ) = −jω
G(xr , x′ )∆κ(x′ )P (x′ , xs )d3 x′ ,
(4.2)
Ω
where
χ(xr ) =



0
1
2

1
xr ∈ Ω
xr ∈ ∂Ω
(4.3)
3
xr ∈ IR \(Ω ∪ ∂Ω).
In equation (4.1) the field inside the scattering domain P (x′ , xs ) can be represented
by a second equation (again based on equation (3.31)) which is given by
Z
P (x′ , xs ) = P i (x′ , xs ) − jω− G(x′ , x′′ )∆κ(x′′ )P (x′′ , xs )d3 x′′ ,
(4.4)
Ω
with x′ ∈ Ω and P (x′′ , xs ) being the total field inside the scattering domain Ω
originating from a source positioned at xs and where ∆κ(xR′′ ) represents the medium
properties inside the scattering domain. The integral sign −Ω means integration over
the domain Ω with a symmetric exclusion for the singular point x′ = x′′ .
In the set of equations (4.1) and (4.4) there are two unknowns. 1) the field inside the
scatterer P (x′ , xs ) and 2) the contrast function ∆κ(x′ ). By substituting equation
Chapter 4: Imaging by double focusing
43
xs
*
∂D
Figure 4.2
D
xr
κ̄ ρ̄
κ ρ̄ Ω
An impulsive point source successively occupies all positions on the surface ∂D surrounding the object. For each source position the scattered field is recorded by receivers
at all positions on the surface.
(4.1) into equation (4.4) the non-linear relation between the measured data and the
contrast function becomes evident.
There are many possible routes to obtain approximate solutions of this non-linear
set of equations. The most well known solution is probably the Born approximation
which linearizes the problem by neglecting the contrast term in equation (4.4). In
this thesis only the solution based on a double focusing technique will be discussed.
4.2
Two-way representation of double focusing; full aperture
To explain the double focusing procedure for two-way wave fields the following experimental set up is chosen; a 3-dimensional scattering domain is illuminated by an
impulsive point source which successively occupies all positions on a surface surrounding the object, see figure 4.2. For each source position the scattered field is
recorded by receivers at all positions on the surface. Note that with this data set all
other possible experiments can be synthesized. The aim is to express the contrasting
medium ∆κ(x′ ) inside the surface in terms of the measured data at the surface and
the Green’s functions.
An expression of the scattered field inside the domain D, in terms of the measurements at the surface ∂D, is obtained by using the focusing integral
I
1
s
{G∗ (x, xr )∂nr P s (xr , xs ) − (∂nr G∗ (x, xr ))P s (xr , xs )} d2 xr .
Φ (x, xs ) =
jω ρ̄ ∂D
(4.5)
Note that equation (4.5) has a similar structure as the backward propagating
Kirchhoff-Helmholtz integral of equation (3.33). Physically Φs (x, xs ) can be interpreted as a wave field obtained by backpropagating the scattered field P s (xr , xs )
44
4.2 Two-way representation of double focusing; full aperture
from the receiver locations xr into the scattering region using the Greens function
G∗ (x, xr ). An alternative interpretation is that it corresponds to focusing the receiver array (common shot gather) at an arbitrary image point inside the surface
∂D using the Kirchhoff integral, Oristaglio (1989). Backpropagating the data from
a receiver array as an intermediate step to imaging or inversion is well known in
seismic migration (Claerbout, 1971; Schneider, 1978; Berkhout, 1982) and diffraction tomography (Devaney, 1982). Note that P s (xr , xs ) represents the scattered
field measured at xr due to a source at xs .
With multiple sources there is a scattered field P s (xr , xs ) at the receiver position
xr for each source position xs . This additional degree of freedom can be used to
derive another expression of the scattered field related to the source dependence of
the scattered field.
I
1
{G∗ (x, xs )∂ns P s (xr , xs ) − (∂ns G∗ (x, xs ))P s (xr , xs )} d2 xs .
Φs (x, xr ) =
jω ρ̄ ∂D
(4.6)
s
Physically Φ (x, xr ) can be interpreted as an image wave field obtained by backpropagating the scattered field P s (xr , xs ) caused by sources at locations xs into the
scattering region using the Greens function G∗ (x, xs ). An alternative interpretation
is that it corresponds to focusing the source array (common receiver gather) at an
arbitrary image point inside the surface ∂D using the Kirchhoff integral.
Substituting the result of equation (4.5) into equation (4.6), by replacing P s (xr , xs )
with Φs (x, xs ), combines the focusing steps for both the source and receiver arrays
and gives
I
1
{G∗ (x, xs )∂ns Φs (x, xs ) − (∂ns G∗ (x, xs ))Φs (x, xs )} d2 xs , (4.7)
Φs (x, x) =
jω ρ̄ ∂D
with Φs (x, xs ) given by equation (4.5). This last equation can be interpreted as
follows; first the image field Φs (x, xs ) is formed by focusing the receiver array at
a point x in the interior of ∂D by integrating over xr ; second, the source array is
focused on the same point by an integral over the source positions xs on ∂D.
Substituting the representation for the scattered field of equation (4.1) in the backward propagating Kirchhoff integral of equation (4.5) gives an expression for the
single focused scattered field in terms of the Greens function, the contrast function
and the scattered field measured at the surface.
Z
I
1
Φs (x, xs ) =
(∂nr G∗ (x, xr ))
G(xr , x′ )∆κ(x′ )P (x′ , xs )d3 x′ d2 xr −
ρ̄ ∂D
Z Ω
I
1
∗
r
G(xr , x′ )∆κ(x′ )P (x′ , xs )d3 x′ d2 xr . (4.8)
G (x, xr )∂n
ρ̄ ∂D
Ω
Interchanging volume and surface integrals and making use of the fact that ∂nr works
Chapter 4: Imaging by double focusing
only on xr gives (after Esmersoy and Oristaglio, 1988),
Z
s
Φ (x, xs ) =
Er (x, x′ )∆κ(x′ )P (x′ , xs )d3 x′ ,
45
(4.9)
Ω
with
1
Er (x, x ) =
ρ̄
′
I
∂D
{(∂nr G∗ (x, xr ))G(xr , x′ ) − G∗ (x, xr )∂nr G(xr , x′ )} d2 xr . (4.10)
From equations (4.9) and (4.2) it is observed that the scattered field and the extrapolated field have a similar integral representation. The difference is that the kernel
Er in equation (4.9) replaces the Green’s function in equation (4.1). The kernel Er
can be interpreted as a Green’s function for the extrapolated field. It relates the
contribution of the secondary sources ∆κ(x′ )P (x′ , xs ) at x′ to the field extrapolated
back to x. The field at x is the superposition of backpropagated waves measured at
the surface ∂D. For a full acquisition geometry Er reduces to a so called ’homogeneous’ Green’s function (Oristaglio, 1989) which is characterized by both incoming
and outcoming (causal and non-causal) waves at infinity and is given by
Er (x, x′ ) = − jω {G(x, x′ ) − G∗ (x, x′ )}
= − jωGh (x, x′ )
and is obtained by using Green’s theorem and the two-way wave equations for the
Green’s functions G∗ (x, xr ) and G(xr , x′ ). An important property of the homogeneous Green’s function Gh (x, x′ ) is that −jωGh (x, x′ ) evaluated at t = 0, thus
after integration over all frequencies, yields a spatial delta function (see appendix in
Oristaglio, 1989).
Substituting equation (4.9) into the double focusing result of equation (4.7) gives
Z
I
1
G∗ (x, xs )∂ns
Er (x, x′ )∆κ(x′ )P (x′ , xs )d3 x′ d2 xs −
Φs (x, x) =
jω ρ̄ ∂D
Ω
Z
I
1
s ∗
(∂n G (x, xs ))
Er (x, x′ )∆κ(x′ )P (x′ , xs )d3 x′ d2 xs .
jω ρ̄ ∂D
Ω
(4.11)
Interchanging volume and surface integrals and making use of the fact that ∂ns works
only on xs gives
Z
s
Φ (x, x) =
Er (x, x′ )∆κ(x′ )Es (x′ , x)d3 x′ ,
(4.12)
Ω
with
Es (x′ , x) =
1
jω ρ̄
I
∂D
{G∗ (x, xs )∂ns P (x′ , xs ) − (∂ns G∗ (x, xs ))P (x′ , xs )} d2 xs .
(4.13)
46
4.3 Two-way representation of double focusing; seismic aperture
Equation (4.12) is the double focusing result and represents an integral over the
scattering domain. This expression can be used to derive an inversion formula for
the scattering object ∆κ(x′ ) at all positions within the domain enclosed by the
surface occupied by the sources and receivers.
The Born approximation for the scattered wave field in equation (4.12) is obtained
by replacing P (x′ , xs ) in equation (4.13) by P i (x′ , xs ) = G(x′ , xs )S(xs ). Note that
for the Born approximation equation (4.13) becomes equal to equation (4.10) (apart
from a scaling factor S), if the sources and receivers occupy the same positions at
the boundary ∂D
I
1
S(xs ) {G∗ (x, xs )∂ns G(x′ , xs ) − (∂ns G∗ (x, xs ))G(x′ , xs )} d2 xs
Es (x′ , x) =
jω ρ̄ ∂D
Es (x′ , x) = Gh (x′ , x)S(xs ).
(4.14)
Within the Born approximation it is possible to arrive at an explicit expression for
the contrast ∆κ(x′ ) function by using the property of the Kernels Er and Es that
in the time domain their time derivative evaluated at t = 0 equals a spatial delta
function. A detailed inspection of the kernels Es and Er is beyond the scope of this
thesis but can be found in Esmersoy and Oristaglio (1988); Oristaglio (1989).
4.3
Two-way representation of double focusing; seismic aperture
Using the backward propagating Kirchhoff-Helmholtz integral of equation (3.29) an
expression of the scattered field inside the domain D, in terms of the measurements
at the surface ∂D and a volume integral over the contrasting domain, is obtained
I
1
s
P (x, xs ) =
{G∗ (x, xr )∂nr P s (xr , xs ) − (∂nr G∗ (x, xr ))P s (xr , xs )} d2 xr
jω ρ̄ ∂D
Z
G∗ (x, x′ )∆κ(x′ )P (x′ )d3 x′ .
(4.15)
−jω
Ω
Note that in comparison with equation (4.5) there is an additional volume integral
which represents the scattering due to the contrasting domain Ω inside D. The
backpropagated field represented by equation (4.15) is now expressed in the true
pressure of the scattered wave field and therefore the symbol P s can be used again.
The contribution from the volume integral over Ω can be neglected if the seismic
acquisition geometry shown in figure 4.3 is taken into account. There are three contributions to the backpropagated wave field at x; ➀ represents a causal contribution
from the surface ∂D1 ; ➁ is a contribution from the surface ∂D0 which is non-causal;
➂ is the contribution from the scattering domain Ω which is also non-causal. For
the configuration in figure 4.3 it can be argued that these non-causal contributions
will cancel each other (except for the evanescent wavefield) and thus can be left out
in the back propagated result.
Chapter 4: Imaging by double focusing
∂D1
47
xs
xr
➀
x
➂
Ω
➁
∂D0
Figure 4.3
Focusing of the receiver array for a seismic acquisition aperture; the sources and receivers are positioned on the surface ∂D1 . The backpropagating contributions from
the surface ∂D0 and the scattering volume Ω are non-causal which leaves only causal
contribution of the surface integral over ∂D1 .
By using this approximation in equation (4.15) it reduces to
Z
1
{G∗ (x, xr )∂nr P s (xr , xs ) − (∂nr G∗ (x, xr ))P s (xr , xs )} d2 xr ,
P s (x, xs ) =
jω ρ̄ ∂D1
(4.16)
which is similar to equation (4.5). Analogous to the previous section (but now with
Φs replaced by P s ) the following expression of the scattered field is obtained
Z
Er (x, x′ )∆κ(x′ )Es (x′ , x)d3 x′ .
(4.17)
P s (x, x) =
Ω
which is similar to equation (4.12). Note that Er in equation (4.17) is represented by
equation (4.10) and Es by equation (4.13), but now with the surface of integration
replaced by ∂D1 . Note also that due to the limited seismic aperture it is not possible
to replace the kernels Er and Es with the homogeneous Green’s function.
Due to a limited aperture of the receiver array at ∂D1 the focused field P s (x, xs )
in equation (4.16) and the scattered field of equation (4.2) can differ significantly.
In this sense the focused field represents a good estimate of the full scattered field
that can be obtained from the observed scattered field at the surface by a linear
operation (backpropagation) that uses no prior knowledge about the nature of the
scattering object.
4.4
One-way representation of double focusing
The double focusing process can also be expressed in the one-way representation
formulation of Berkhout (1996a, 1996b). The advantage of the one-way formulation
4.4 One-way representation of double focusing
∆
*
∆
xs
xr
∆
∆
∆
∆
∆
∂D1
∆
48
x
Ω
∂Ω
∂D0
Figure 4.4
Focusing of the source and receiver arrays to a point in the subsurface. Note that if the
halfspace above (and including) ∂D1 is taken homogeneous and isotropic then the field
measured at ∂D1 consists of only upgoing waves.
is that it clearly shows the relation between the result after the first focusing step
and the used macro model to construct the backpropagating Green’s function. At
the end of this section the double focusing procedure is also shown in the more
convenient matrix notation.
To explain the double focusing procedure for one-way wave fields the following experimental set up is chosen; a 3-dimensional scattering object is illuminated by an
impulsive point source which successively occupies all positions on two planar surfaces above and below the object (∂D1 and ∂D0 , see figure 4.4). For each source
position the scattered field is recorded by receivers at all positions on the surface.
The aim is to express the scattering operator Θ inside the surface in terms of the
measured data at the surface ∂D1 alone.
4.4.1
Integral representation
To derive the integral representation of the double focusing procedure for one-way
wave fields the following Green’s state is chosen: the propagation operator Λ̄ = Λ
is chosen the same as in the actual medium and the scattering operator for the
Green’s state is equal to zero, Θ = O. With these choices the Green’s matrix
G(x, x′ ) contains only diagonal elements which are given by equation (3.44). Note
that the source is positioned outside domain D and the scattering domain is chosen
inside domain D.
Backward propagation of the measurements at the detector position to a point x
inside domain D can be done by using the backpropagating surface integral representation of equation (3.41). The backward propagated result with the detector
Chapter 4: Imaging by double focusing
positions at the surface ∂D is then given by
Z
KG∗ (x, xr )KP s (xr , xs )n3 d2 xr +
P s (x, xs ) = −
Z∂D
KG∗ (x, x′ )K(Θ(x′ ) − Θ̄(x′ ))P (x′ )d3 x′
49
(4.18)
Ω
Note that there is an extra volume integral in comparison with equation (3.43). This
volume integral is due to the presence of the scattering domain Ω inside domain D.
Taking the configuration shown in figure 4.4 into account there is only an upgoing
scattered wave field at the surface ∂D1 . Suppose that, unlike in figure 4.4, the surface
∂D0 is positioned above the scattering domain then the volume integral vanishes and
the surface integral over ∂D0 can be neglected, again ignoring evanescent waves (see
also the discussion after equation (3.43)) which leaves only the surface integral over
∂D1 . If the surface ∂D0 is positioned below the scattering domain (as in figure 4.4)
it is evident that the contribution from the surface ∂D1 is not changed which means
that the contributions from the volume integral and the surface integral over ∂D0
cancel each other. Taking these arguments into account gives for the backpropagated
upgoing scattered field
Z
−,s
Wp+,∗ (x, xr )P −,s (xr , xs )d2 xr ,
(4.19)
P (x, xs ) =
∂D1
where, according to equation (3.44) Wp+ (x, xr ) = G+,+ (x, xr ) and where
P −,s (x, xs ) represents the scattered wave field at x after focusing of the receiver
array.
Making use of the redundancy present in the data the same focusing procedure used
in equation (3.43) can also be used for the surface ∂D1 at the positions where the
sources are placed. This focusing step along the source surface is represented by
Z
Wp+,∗ (x, xs )P −,s (xr , xs )d2 xs ,
(4.20)
P −,s (x, xr ) =
∂D1
where P −,s (x, xr ) represents the scattered wave field at x after focusing of the shot
array (common receiver gather).
The double focusing result for one-way wave fields is obtained by substituting the
result of the focused receiver array of equation (4.19) into the focused source array of equation (4.20) by replacing xr by x in P −,s (xr , xs ) using Wp+,∗ (x, xs ) =
Wp−,∗ (xs , x), gives
Z
−,s
P −,s (x, xs )Wp−,∗ (xs , x)d2 xs
P (x, x) =
∂D1
Z
Z
=
Wp+,∗ (x, xr )P −,s (xr , xs )Wp−,∗ (xs , x)d2 xr d2 xs . (4.21)
∂D1
∂D1
50
4.4 One-way representation of double focusing
Note that both the source and receiver array are focused on the same point x in
space. Equation (4.21) shows that the double focusing procedure is carried out by
integrating the measured data over the source and receiver surface towards a point
in the subsurface.
The scattered field P −,s (xr , xs ) occuring in equation (4.21) can be represented by
the volume integral of equation (3.49)
P −,s (xr , xs ) =
Z
Ω
Wp− (xr , x′ )R+ (x′ )Wp+ (x′ , xs )S0+ (xs )d3 x′ .
(4.22)
Note that in this equation R is a reflection operator defined by equation (3.48).
Substituting this W RW representation of the scattered wave field into equation
(4.21) gives after some manipulations
P −,s (x, x) =
Z
Ω
Er− (x, x′ )R+ (x′ )Es+ (x′ , x)d3 x′ ,
(4.23)
where it is assumed that the scattering domain Ω and the focus point x are both
positioned below the surface ∂D1 . Er− and Es+ are then given by
Er− (x, x′ )
=
Es+ (x′ , x) =
Z
∂D1
Z
∂D1
Wp+,∗ (x, xr )Wp− (xr , x′ )d2 xr
(4.24)
Wp+ (x′ , xs )S0+ (xs )Wp−,∗ (xs , x)d2 xs
(4.25)
Making use of the integral expression for the reflection operator and writing the
surface integral explicitly for the horizontal plane gives
Z Z
−
Er (x, x′ )R+ (x′ , x′′ )Es+ (x′′ , x) x′′ =x′ d2 x′′ d3 x′
P −,s (x, x) =
(4.26)
Ω
IR2
3
3
An interpretation of equations (4.24), (4.25) and (4.26) is shown in figure 4.5. Er−
in equation (4.26) can be interpreted as forward propagation from a reflection point
at x′ to the receiver array at the surface followed by backward propagation from
the receiver array to a point x in the subsurface. Note that by choosing x = x′ the
receiver array is focused on a point on the reflector. This means that the scattered
data measured by the receivers at ∂D1 is used to construct the scattered field at x
originating from a source positioned at xs . Es+ represents backward propagation of
the contributions of the different sources at the reflection point x′ to a point x in
the subsurface.
Equation (4.26) can be simplified further by making use of the fact that the complex
conjugate of the propagation operators Wp are an approximation of the inverse
propagation operators (see the discussion after equation (3.43)). Using this property
Chapter 4: Imaging by double focusing
xr
xs
∆
∂D1
51
*
Es+ (x′′ , x)
Er− (x, x′ )
x
x3
x′ x′′
Ω
∂D0
Figure 4.5
Focusing of the source and receiver arrays to a point in the subsurface. Note that the
interior of both the source and receiver domains must have an intersection in which the
double focusing can be carried out.
equations (4.24) and (4.25) can be interpreted as follows;
Z
Wp+,∗ (x, xr )Wp− (xr , x′ )d2 xr
Er− (x, x′ ) =
∂D1
= δ(xα − x′α )
if x′3 = x3
Z
Wp+ (x′′ , xs )S0+ (xs )Wp− (xs , x)d2 xs
Es+ (x′′ , x) =
(4.27)
∂D1
= δ(x′′α − xα )S0+
if x′3 = x3
(4.28)
where the assumption is made that the downgoing source function is independent
of the source position; S0+ (xs ) = S0+ . So, the surface integral can be represented by
the delta function if the depth of the focus point is chosen equal to the depth of the
reflector. Which is equivalent with the statement that the back transformation of
Er and Es to the time domain gives a spatial delta function at t = 0 (if x′3 = x3 )
(note the resemblance with the two-way expressions). The non-causal times (t < 0)
can be interpreted as the influence of the scattering domain above the focus point
(if x′3 < x3 ) and the causal times (t > 0) are the contributions from the scattering
domain below the focus point (if x′3 > x3 ). Substituting these results into equation
(4.26) gives the following representation of the reflection operator
P −,s (x, x) = R+ (x, x)S0+ + ε1 + ε2 ,
(4.29)
where ε1 represents the non-causal contribution and ε2 the causal contribution with
respect to the focus point. Equation (4.29) shows that the double focusing procedure
gives a recovery of the reflection operator at t = 0. The obtained reflectivity information is constrained by the strength of the reflection, the aperture of the receiver,
the coverage of the sources and the bandwidth of the data. The double focusing procedure can be interpreted as a consistent way of integrating the redundancy present
in the ensemble of all the data.
52
4.4 One-way representation of double focusing
4.4.2
Matrix representation
The focusing process for one-way wave propagation can be expressed in the WRW
matrix formulation which was introduced in section (3.4). The derived forward
model of seismic reflection data, backscattered from one depth level at zm , is given
by
P(zr , zs ) = D− (zr )W− (zr , zm )R+ (zm )W+ (zm , zs )D+ (zs )S(zs ),
(4.30)
where zr represents the receiver level, zm the reflection level and zs the source
level. Focusing in emission can be regarded as a weighted summation (in phase and
amplitude) along the common receiver arrays in such a way that the constructed wave
front originates from a notional source at a point in the subsurface. The weighting
operator used in this process is also called the focusing or synthesis operator, because
the operator synthesizes the response of a focusing areal source from the seismic data.
The principle of combining shot gathers at the surface for the synthesis of areal source
responses, also referred to as areal shot record technology, was already introduced by
Berkhout (1992) for controlled illumination in prestack depth migration. Rietveld
(1995) has shown many examples for different illuminating areal sources. Note
that one row of the matrix P(zr , zs ) represents one common receiver array and
one column represents one common shot gather (see also Appendix B). Thus the
synthesis operator for focusing in emission works on the rows of the matrix P(zr , zs )
and a synthesis operator for focusing in detection works on the columns of the matrix
P(zr , zs ).
The focusing operator is defined with respect to the coordinates of a common receiver
or common source array. The focusing operator for the receiver array, applied to the
left side of the right-hand side of equation (4.30), is defined as
Fi− (zm , zr )D− (zr )W− (zr , zm ) = Ii− (zm )
(4.31)
∗ −1
Fi− (zm , zr ) ≈ Ii− (zm ) W+ (zm , zr ) D− (zr )
with Ii− (zm ) a unit row vector with a 1 at the ith position at depth zm and Fi− (zm , zr )
the focusing operator acting at the receiver positions at the surface. Note that
the approximation in equation (4.31) refers to the approximation of the inverse
−1
of the propagation operator W− (zr , zm ) by its matched filter [W− (zr , zm )]
≈
∗
[W+ (zm , zr )] .
The focusing operator for the source array, applied to the right side of the right-hand
side of equation (4.30), is defined as
Ij+ (zm ) = W+ (zm , zs )D+ (zs )S(zs )Fj+ (zs , zm )
−1
∗
D+ (zs )S+ (zs )
W− (zs , zm ) Ij+ (zm ) ≈ Fj+ (zs , zm )
(4.32)
Chapter 4: Imaging by double focusing
Figure 4.6
focusing in detection
∗
∆
∗
j
∆
focusing in emission
53
i
Focusing in detection positions a virtual receiver on a reflecting boundary and focusing
in emission positions a virtual source on a reflecting boundary.
with Ij+ (zm ) a unit column vector with a 1 at the j th position at depth zm and
Fj+ (zs , zm ) the focusing operator acting at the source positions. The focusing operators F ± perform a summation along the receiver positions (F − ) in a common
shot gather or a summation along the source positions (F + ) in a common receiver
gather. This summation (or synthesis) is carried out for all source and receiver
positions available. The focusing operator for detection is given by the ith row of
[W+ ]∗ [D− ]−1 and the focusing operator for emission is given by the j th column of
[D+ S+ ]−1 [W− ]∗ .
Substituting equation (4.31) into equation (4.30) gives an expression of the data
after focusing of the detector array
Fi− (zm , zr )P(zr , zs ) = Pi− (zm , zs ) = Ii− (zm )R+ (zm )W+ (zm , zs )D+ (zs )S(zs ) (4.33)
where equation (4.33) is an expression for the so called Common Focus Point (CFP)
gather for focusing in detection and is shown in figure 4.6.
Note that if the backpropagating Green’s function, which is expressed by Fi− (zm , zr ),
represents correct propagation then the result of the first focusing step Pi− (zm , zs ),
represented by equation (4.33), is in traveltime equal to the time reversed focusing
operator Fj+ (zs , zm ), represented by equation (4.32). This is a very important property of the first focusing step and is called the principle of equal traveltime. This
principle plays a fundamental role in the updating procedure described in chapter 5
where a seismic image can be built up without knowing the background model. The
principle can also be used to update the initial background model of the Green’s
function. Note that for a CFP gather designed for focusing in detection one should
compare the traveltimes for an operator defined at the source position, which is the
operator for focusing in emission.
Substituting equation (4.32) into equation (4.30) gives an expression for the focusing
of the source array
P(zr , zs )Fj+ (zs , zm ) = Pj− (zr , zm ) = D− (zr )W− (zr , zm )R+ (zm )Ij+ (zm )
(4.34)
54
4.4 One-way representation of double focusing
double focusing
∆
∗
ij
Figure 4.7
Focusing in detection and emission positions a virtual source and a virtual receiver on a
reflecting boundary. In confocal imaging i=j.
where equation (4.34) is an expression for the Common Focus Point (CFP) gather
for focusing in emission and is shown in figure 4.6.
Focusing of both the detector and the source array by combining equation (4.31)
and equation (4.32) into equation (4.30) gives
Pij− (zm ) =Fi− (zm , zr )P(zr , zs )Fj+ (zs , zm )
=Ii− (zm )R+ (zm )Ij+ (zm )
+
=Rij
(zm )
(4.35)
which is the double focusing result shown in figure 4.7 for one reflecting depth level.
Taking into account the presence of the out-of-focus responses from the other depth
levels as well, equation (4.35) need be replaced by
+
Pij− (zm ) =Rij
(zm ) + ε1ij + ε2ij
(4.36)
where ε1ij represents the contribution from above and ε2ij from below the focusing
level. Note that this result is the same as equation (4.29) when i = j.
Chapter 5
CFP technology
For those readers who have skipped the theoretical chapters, this chapter is a good
starting point to get an understanding of the possibilities of the Common Focus Point
(CFP) technology. The schemes and expressions, which were derived in the previous
chapters, are in this chapter explained by using simple numerical experiments which
show clearly the fundamental principles of the double focusing procedure. At the
end of this chapter it will be made clear that the double focusing procedure is a
good starting point for the analysis and imaging of seismic data.
The interpretation and understanding of CFP gathers has led to significantly new
insights and new processing schemes. In section 5.2 the construction of a CFP gather
and the applications of the CFP technology is explained and illustrated with simple
numerical examples. In section 5.3 the second focusing step is explained and the
CFP imaging technique is discussed. In the last sections, resolution and amplitude
analysis (5.4), the construction of 3-dimensional CFP gathers (5.5) and the latest
developments (5.6) are discussed and illustrated.
5.1
Areal shot record technology
The common focus point technology has its origin in the areal shot record technology
which was proposed by Berkhout (1992) and developed within the DELPHI project
by Walter Rietveld (1995). An areal shot record represents the response of the subsurface due to a source wave field with lateral extent. Such an areal shot record
can be constructed by integration of weighted common receiver gathers followed by
combining the integration results into one gather. The weighting of the individual
sources in the common receiver gathers is determined by a so-called synthesis operator which is calculated by a forward modeling step in the macro model. The
synthesis operator is designed in such a way that the illumination of the desired target takes place. For example an illuminating wave field perpendicular to a reflector
at the target (normal incidence) is very useful to determine the macro boundaries
56
5.1 Areal shot record technology
depth [m]
-1500
0
-750
lateral position [m]
0
750
1500
500
1000
a. interference pattern of 3 sources
depth [m]
0
500
1000
b. interference pattern of 7 sources
depth [m]
0
500
1000
c. interference pattern of 21 sources
depth [m]
0
500
1000
Figure 5.1
d. interference pattern of 201 sources
Huygens-Fresnel construction for a wavefront in the subsurface by using a different number of sources at the surface. Note that in figure d), where 201 sources are used to construct the wavefront, only a small part of the individual wavefronts, visible in a,b and c,
(the Fresnel zone) give a constructive contribution to the total wavefront.
Chapter 5: CFP technology
57
in the area of interest. In general every desired wave field in the subsurface can be
built up by using weighted sources at the surface. This so called Huygens principle
is shown in figure 5.1 for 2-dimensional point sources (which are equivalent to 3dimensional line sources) at the surface and an areal wavefront in the subsurface.
From figure 5.1 it can also be seen that due to the interference of the different circular
wavefronts, originating from the individual sources at the surface, only a small part
of the wavefronts have a contribution to the actual wavefront. This combination of
Huygens construction with the principle of interference is called the Huygens-Fresnel
principle (Born and Wolf, 1970). The basic idea of the Huygens-Fresnel theory is
that the wave field at an observation point can be constructed from the superposition of secondary sources positioned at a surface between the observation point and
the source. In the previous chapter this idea was mathematically expressed in the
Kirchhoff integral of equation (3.40). In this chapter it will be made clear that the
Fresnel zone plays an important role in the construction of CFP gathers.
One of the advantages of controlled illumination is that it is an efficient and accurate
way of performing a pre-stack depth migration. By using only a limited number of
illumination angles at target level a good pre-stack depth image can be obtained.
Examples of the imaging results obtained with controlled illumination on numerical
and real data can be found in Rietveld (1995). A special case of controlled illumination is a point illumination. This point illumination was initially used for macro
model verification, but by recognizing and using the special properties of the point
illumination it became clear that it is an excellent intermediate domain for velocity
and AVO analysis as well. The areal source wave field defined for a grid-point in the
subsurface is called the Common Focus Point gather.
5.2
First focusing step
The Common Focus Point (CFP) gather is constructed from seismic data acquired
within a pre-specified source and detector aperture. To construct the CFP-gather
from the data an initial synthesis (also called focusing) operator is needed. This
initial synthesis operator can be based on stacking velocities or an initial macro
model. The initial synthesis operator is calculated by positioning a point source at
the desired grid-point in the subsurface followed by a forward modeling algorithm to
calculate the source response at the surface. Measuring its response at the detector
positions defines an operator for focusing in detection and measuring its response
at the source positions defines an operator for focusing in emission. As shown
by equations (4.31) and (4.32), the time reverse (complex conjugate in frequency
domain) of the forward modeled response defines the synthesis operator. In figure
5.2, a synthesis operator is shown for a grid-point defined in the middle of a synclinal
model, at the synclinal reflector. The forward modeling scheme used to model
the response is a 2-dimensional recursive depth extrapolation based on weighted
58
5.2 First focusing step
depth [m]
one-way time [s]
-1500
0
-750
lateral position [m]
0
750
-1500
-1.0
1500
0.5
-0.5
1.0
0
0
0
500
500
1000
1000
1500
1500
a. forward modeling of focusing operator
Figure 5.2
-750
lateral position [m]
0
750
1500
b. focusing operator
Modeling of the focusing operator in the synclinal model. Note that the time-reversed
focusing operator is equal to the forward modeling result.
least-square operators as described in appendix A. Figure 5.2a shows the forward
modeling result and figure 5.2b shows the focusing operator. If the forward operator
is calculated in the correct model the focusing operator will have its focus on the
defined grid-point. Note that the vertical axis along the operator is the one-way
time axis.
To calculate one trace out of a CFP gather for focusing in detection the synthesis
operator measured at the detector positions must be convolved in time with a shot
record followed by a summation along all traces in the record. This procedure
is also shown by the integral of equation (4.19) and the matrix multiplication of
equation (4.31). Performing this procedure, with the same operator for all shot
records available defines the CFP gather for focusing in detection. As mentioned
before not all traces in the shot record have an equal contribution to the CFP trace,
only that part of the CFP corrected shot record (the result after convolution with
the operator but before summation of the traces) which lies in the Fresnel zone
contributes to the CFP trace. Note that the assumption is made that the high
frequency approximation is valid such that it is allowed to speak of a Fresnel zone.
The Fresnel zone represents that part of the shot record which is related to the chosen
focus point and is therefore dependent on the focusing operator. The acquisition
geometry of the shot record also plays a role because the measured data should
at least contain reflection information from the defined focus point. It is therefore
important to discuss first the Huygens-Fresnel principle before the first focusing step
is explained in more detail.
Chapter 5: CFP technology
59
S
xr
a
xs
Figure 5.3
5.2.1
xo
b
x
Pictorial representation of the Fresnel zone. Due to the interference of the secondary
sources on the wavefront S only a small part of S, the Fresnel zone, contributes to the
wave field at x. The contribution depends on the path difference between x and a point
on the wavefront S, xr .
Huygens-Fresnel principle
There are many ways to introduce the Fresnel zone. In this thesis a definition close
to the original definition of Fresnel is used. In figure 5.3 a circular wavefront S is
shown which originates from a source at xs and with x the point where the wave
field is to be determined. In accordance with the Huygens-Fresnel principle each
element on the wavefront S is regarded as the center of a secondary point source.
Now let b = |x − xo | the shortest distance between the point of observation and the
wavefront S. Then by assuming a homogeneous medium in-between xs and x the
most important contribution from the wavefront at x comes from the point at xo .
The contribution of the point xr on the wavefront to x is determined by the distance
a = |x−xr |. This contribution is constructive if the path difference a−b = k ∗λ with
k a positive integer, and λ the wavelength. Destructive interference occurs when the
path difference a − b = (2k + 1) ∗ λ2 .
Using stationary phase analysis (Bleistein, 1984) this interference behavior can be
made more clear in a mathematical sense. The contribution from all secondary
sources to x is represented by a summation along the wavefront given by the integral
Z
P (x, xs ) =
P (xr , xs )Ag (x, xr ) exp (jΦg (x, xr ))d2 xr ,
(5.1)
S
where Ag and Φg represent respectively the amplitude and phase of the Green’s
function traveling from a point on the wavefront S to the point of observation
x. As the point xr runs along the surface of integration, the function Φf =
amplitude
60
5.2 First focusing step
1.2
1.2
1.0
1.0
0.8
0.8
0.6
0.6
0.4
0.4
0.2
frequency = 60 Hz (af = 165 m)
0
500
1000
integration interval [m]
1500
0.2
0
frequency = 5 Hz (af = 575 m)
0
500
1000
integration interval [m]
1500
freq
uen
cy
amplitude cross section at image time
1.2
amplitude
[1/s
]
in
te
gr
at
io
n
in
te
rv
al
[m
]
amplitude
0
1.0
0.8
0.6
0.4
0.2
0
Figure 5.4
0
500
1000
integration interval [m]
1500
The two top pictures show the amplitude of the integral in equation (5.1) as function
of the integration interval for two different frequencies. The kernel consists of the 2dimensional Green’s function for a dipole source in a homogeneous medium with c =
2000 m/s and a depth of 800 m. The 3-dimensional picture in the middle shows the
same amplitude for a range of frequencies and integration intervals.
Φp (xr , xs ) + Φg (x, xr ), where Φp (xr , xs ) is the phase of the wavefront P (xr , xs ) at
S, will change very many cycles, so that both the real and imaginary parts of the
integrand will change sign many times. As a consequence the contributions from
the various elements will in general virtually cancel each other out (destructive in-
Chapter 5: CFP technology
61
terference). The situation is, however, different for an element which surrounds a
point (called critical point or pole, in figure 5.3 the point xo ) where Φf is stationary. Here the integrand varies much more slowly and may be expected to give a
significant contribution. Hence, when the wavelength is sufficient small, the value of
the integral is determined substantially by the behavior of Φf in the neighborhood
of points where Φf is stationary.
The area of constructive interference surrounding the stationary phase point is called
the Fresnel zone. The limit of constructive interference is defined differently by
various authors in terms of fractions of the wavelength λ. Note, however that the
Fresnel zone is a weighting function with no sharp cutq
off point (Lindsey, 1989).
In this thesis the definition of Berkhout (1984), af = 2 (z + λ8 )2 − z 2 is used for
convenience. By choosing in equation (5.1) P (xr , xs ) = 1; a plane wave with zero
phase on the flat surface S, the interference behavior around the Fresnel zone can
be analyzed by calculating the amplitude of the integral. This amplitude should
be equal to 1 to get a correct amplitude in the integration result. In figure 5.4
the amplitude of the integral in equation (5.1) is shown for different integration
intervals and frequencies. The kernel consists of the 2-dimensional Green’s function
for a dipole source in a homogeneous medium with c = 2000 m/s and a distance of
800 m.
The two top pictures of figure 5.4 show the amplitude as function of the integration
interval for two different frequencies. Note that even for the high frequency (small
λ) a large integration interval must be taken to obtain a correct amplitude in the
integration result. For someone who is interested in true amplitudes this seems a
disappointing result, because it means that very large apertures should be taken into
account to obtain correct amplitudes in the integration result. However, by looking
at the 3-dimensional picture in the middle of figure 5.4 it can be argued that due
to the interference off all frequencies, the integration result in the time domain will
give a better result. In the bottom picture the cross-section at the imaged time
( zc ) is shown for different integration lengths. In this picture it is observed that
an integration length of 200 m, corresponding to the Fresnel zone of the central
frequency of 35 Hz, is sufficient to obtain the correct amplitude in the integration
result (for more details see Rietveld, 1995, appendix B).
5.2.2
Construction of CFP trace: dipping layer
To illustrate the importance of the Fresnel zone in the first focusing step a simple
numerical example is used. In this example a CFP gather is constructed from data
modeled with a fixed receiver spread above a dipping reflector. The focus point is
chosen on the dipping reflector in the middle of the acquisition geometry (see figure
5.5c). The center of the reflector is positioned at 800 m depth and the velocity above
the reflector is 2000 m/s.
62
5.2 First focusing step
Figure 5.5 shows how a CFP trace for focusing in detection is constructed by using
1500
●
0
750
1500
-1500
0
one-way time [s]
-750
-750
0
750
1500
0.5
1.0
-750
1
0
750
d. time reversed focusing operator
1500
➝
-1500
0
2
-750
0
750
1500
0.5
➝
two-way time [s]
1500
b. shot record with xs = 1250
c. rays for focusing operator
-1500
0
1.0
e. overlay of operator on shot
-750
0
750
f. CFP corrected shot record
1500
-1500
0
one-way time [s]
-1500
0
depth [m]
lateral position [m]
0
750
2
750
750
1500
-750
1
a. rays for shot record
-1500
0
1500
-1500
0
two-way time [s]
lateral position [m]
0
750
750
1500
depth [m]
-750
one-way time [s]
depth [m]
-1500
0
-750
0
750
1500
0.5
1.0
g. path of traveltime after synthesis
Figure 5.5
h. CFP trace
Construction of one CFP trace for focusing in detection given a shot record (b) and a
focusing operator (d). The only contribution in the CFP trace (h) comes from the Fresnel
zone (approximately 350 m) indicated with an arrow in (e) and (f).
Chapter 5: CFP technology
63
a shot record (b) and a focusing operator (d). Time convolution of the synthesis
operator with the shot record gives the CFP corrected shot record shown in (f).
The arrow in the pictures (e) and (f) indicates the Fresnel zone. The lateral position
of the center of the Fresnel zone (the stationary phase point) can be determined
by following the ray-path from the source, via the focus position, to the receiver
position (indicated with an arrow in figure 5.5a). Another way to determine the
stationary phase point can be done by superimposing the times of the operator over
the shot record and shifting the operator times along the vertical axis until a part
of the operator is tangent to the event in the shot record (e).
After summation over all the traces in the CFP corrected shot record (f) the CFP
trace is obtained. The only constructive ’interference’ of the event in (f) occurs at
the Fresnel zone, which is indicated with an arrow. The constructed CFP trace is
positioned in the CFP gather at the lateral position of the source in the shot record
(h). The CFP trace is positioned at the source position because the time of the
event in the CFP trace is equal to the time the wavefield needs to propagate from
the source at the surface to the focus point at the reflector (g). This result is in
agreement with equation (4.33) where it has been shown that one ’leg’ of the two-way
traveltime is compensated by the focusing operator in the first focusing step. By
using more shot records more CFP traces can be constructed with the same operator
and finally when all available shot records are used the CFP gather is completed.
Another way to show how a CFP gather is built up from different shot records
is shown in figure 5.6. The same operator and model configuration of figure 5.5
is used, but now all shots are used to construct the CFP gather (201 shots with a
sampling distance equal to the (minimum) receiver spacing = 15 m) and the number
of receivers is different for the different results shown in figure 5.6. In figure 5.6a
only 7 receiver per shot record are used in the construction of the CFP gather for
focusing in detection. This means that the Fresnel zone as shown in figure 5.5f is
sparsely sampled and constructive interference does not occur after summation of
only 7 traces. By looking carefully in figure 5.6a the 7 contributions of the individual
receivers can still be recognized. Using more receivers in the shot record shows, from
picture (a) to (d), how the Fresnel zone is built up in the CFP domain.
In figure (e) the complete CFP gather is shown with the times of the synthesis operator superimposed on the result. Remembering the principle of equal traveltime,
introduced in the previous chapter; which states that if the focusing operator represents correct propagation in the model then in the time domain the result of the
first focusing step Pi− (zm , zs ), represented by equation (4.33), is in traveltime equal
to the time reversed focusing operator Fi− (zm , zr ), represented by equation (4.32).
However, by looking at the overlay (e) it is observed that in the left part in the
CFP gather, the times of the operator don’t match with the times of the CFP
response. This seems confusing at first sight, but considering the lateral position
64
5.2 First focusing step
one-way time [s]
-1500
0
-750
lateral position [m]
0
750
1500
0.5
-1500
0
1500
1.0
a. 7 receivers per shot
-1500
0
one-way time [s]
lateral position [m]
0
750
0.5
1.0
-750
0
b. 21 receivers per shot
750
1500
0.5
-1500
0
-750
0
750
1500
750
1500
0.5
1.0
1.0
c. 41 receivers per shot
-750
0
d. 201 receivers per shot
750
1500
shadow zone
➝
0.5
1.0
e. overlay of operator and CFP gather
Figure 5.6
depth [m]
-1500
0
one-way time [s]
-750
-1500
0
-750
0
750
1500
f. shadow zone in aperture
Construction of one CFP gather by using all shots, but with a different number of receivers per shot record. Note that due to the limited aperture there is a shadow zone in the
CFP gather (f).
of the Fresnel zone for the shots positioned at the left side it is observed that the
Fresnel zone falls outside the receiver aperture. So, the mismatch in time between
the focusing operator and the CFP response is due to limited aperture. The zone
where the Fresnel zone is ’shifted’ outside the aperture is called the shadow zone.
The ray paths of the shot record with its stationary phase ray just inside the receiver
aperture is shown in (f). The CFP gather shows clearly that the shot records in the
shadow zone don’t contain the full reflection information of the defined focus on the
reflector.
Another thing which can be observed in the CFP gather (d) is that there are two
artifacts present above the CFP response. These artifacts are due to the finite
aperture of the receiver array and originates from the truncation of the synthesized
Chapter 5: CFP technology
65
reflection response in figure 5.5f. The destructive interference of the traces at the
boundaries of the CFP corrected shot record is not complete due to the ’missing’
contribution of the traces outside the aperture. There are several ways to reduce
these artifacts; one method is tapering around the Fresnel zone in the CFP corrected
shot record, another method makes use of an analytical method to remove finite
aperture artifacts. In the next subsection the analytical method is discussed in more
detail because of its elegant solution to the finite aperture problem.
Removal of finite aperture artifacts
In the frequency domain implementation of the synthesis process, the shot record
and the focusing operator are transformed to the frequency domain and multiplied
together. The result of this multiplication is summed over all receivers in the shot
record to produce one trace of the CFP gather. Wapenaar (1991) has shown that
due to the finite aperture of the data, artifacts are introduced in this final result.
These artifacts can be suppressed by using a taper at the edges of the data set.
However, tapering only suppresses the artifacts (and a part of the data as well) but
does not remove the artifacts completely. Therefore Wapenaar (1991) proposed a
new technique (worked out in detail by Timmerman (1993)) which calculates in an
analytical way (using diffraction theory), the artifacts originating from the finite
aperture and subtract these calculated artifacts from the integration result. This
analytical method gives more insight in the influence of the acquisition geometry
in constructing a CFP gather and shows the importance of the Fresnel zone. The
summation over all receivers in a shot gather, to obtain one trace in the CFP gather,
is represented by the following integral (which is the same as equation (5.1))
Z ∞
I=
f (x)ejφ(x) dx,
(5.2)
−∞
where f (x) and φ(x) represent the amplitude and phase, respectively. Due to the
finite acquisition aperture, equation (5.2.2) is approximated by
Z xb
f (x)ejφ(x) dx,
(5.3)
< I >=
xa
where xa and xb are the begin- and end-points of the acquisition aperture. These
finite integration limits cause artifacts. Wapenaar expresses the correct result I in
terms of the approximated result < I > plus two correction terms, as follows:
I =< I > +Ia + Ib
where
Ia =
Z
xa
Z
∞
(5.4)
f (x)ejφ(x) dx
(5.5)
f (x)ejφ(x) dx.
(5.6)
−∞
and
Ib =
xb
66
5.2 First focusing step
1500
lateral position [m]
0
750
1500
-1500
0
one-way time [s]
-750
750
●
-750
lateral position [m]
0
750
1500
0.5
1.0
a. subsurface model
-1500
0
one-way time [s]
depth [m]
-1500
0
-750
0
0.5
b. CFP gather
750
1500
-1500
0
-750
0
750
1500
0.5
1.0
1.0
c. calculated artifacts
Figure 5.7
d. CFP gather without artifacts
Finite aperture artifacts removal in the CFP gather. The sharp cutoff in the artifacts
free CFP response at x = −330 m represents the start of the shadow zone and is given
by the shot record indicated in figure 5.6f. Note that all figures are plotted on the same
scale.
By adding Ia and Ib to the finite integration result the finite aperture artifacts
are fully compensated. The method of Wapenaar makes a first and second order
approximation to the finite aperture artifacts described in equations (5.5) and (5.6).
The CFP gather in figure 5.7b is shown for the same dipping reflector example which
was discussed earlier. Note that the artifact which starts on the upper left side of
figure 5.7b disturbs the CFP response at the right side. The calculated artifacts,
based on the second order approximation, are shown in figure 5.7c. Subtraction of
these artifacts from figure 5.7b gives the artifacts free result of figure 5.7d. Note the
almost perfect removal of the ’shadow zone’ present in figure 5.7b. The remaining
part of the artifact is approximately 5% of the amplitude of the main event. The
sharp cutoff in the artifacts-free CFP response at x = −330 m represents the start
of the shadow zone and is given by the shot record indicated in figure 5.6f.
The proposed method breaks down if more than one event is present in the integration Kernel, in that case the phase cannot be determined uniquely for a single event.
This limitation of the method can be taken into account by selecting one event in
a time window around the times given by the CFP operator. However, removal of
the finite aperture artifacts is not essential in further processing steps because they
don’t disturb image quality or the amplitude response in the CFP gather.
one-way time [s]
Chapter 5: CFP technology
-1500
0
-750
0
750
1500
0.5
1.0
0
0
-1500
0
0.5
0.5
1.0
1.0
one-way time [s]
a. synt reflector depth = 800 m
-1500
1.0
-750
0
750
1500
1.5
2.0
1.0
0
750
1500
0
0
0.5
1.0
b. with noise (S/N = 5 dB)
0
-1500
1.0
1.5
1.5
2.0
2.0
c. no noise reflector depth = 3000 m
Figure 5.8
-750
67
-750
0
750
1500
1.0
0
1.5
2.0
d. with noise (S/N = 5 dB)
Signal enhancement due to the constructive summation in the CFP corrected shot record.
For a deeper reflector (b) the Fresnel zone is larger and hence the signal to noise ratio in
the CFP gather will be better (right trace) than the ratio in one single trace (left trace).
The influence of noise
Noise in the data will in general distort the quality of the CFP gather, however,
due to the summation of the events in the Fresnel zone the signal to noise ratio
in the CFP gather is improved in comparison with the signal to noise ratio in the
shot records the CFP gather is built from. This is illustrated in the CFP corrected
shot record of figure 5.8a where ’random’ noise is added to the shot record, shown
in figure 5.5b. Also for a deeper reflector, which has an even worser S/N ratio and
shown in figure 5.8d, the Fresnel zone is larger and the signal to noise ratio of the
CFP trace shown in figure 5.8d is better than the signal to noise ratio of a single
trace in the shot record (left trace).
5.2.3
Construction of CFP gathers: synclinal model
The construction of a CFP gather in a more complicated model is explained by
using numerical data based on the model shown in figure 5.9a. The numerical data
is modeled with a fixed acquisition spread where the source positions are defined
at every receiver position (201 shot positions with ∆x = 15 m). The source has
a dipole character and its signature is given by a Ricker wavelet with a frequency
peak at 26.4 Hz. For the forward modeling of the data an acoustic finite difference
algorithm is used. The subsurface model includes a diffraction point at z = 1000,
x = −750 m and a negative reflection coefficient for the wedge located in the right
corner of the model. In the remainder of this chapter this syncline model will be
used to explain the different processes in the double focusing procedure.
The synthesis process for a focusing receiver with a focus point defined at the synclinal interface at x = 0 and z = 950 m (the focus point is indicated with a black
bullet in figure 5.9a) is shown in detail in figure 5.9. The time reversed focusing
68
5.2 First focusing step
-1500
0
-750
lateral position [m]
0
750
1500
-1500
0
-750
lateral position [m]
0
750
1500
c=2200
1000
c=2700
●
c=2400
c=3000
1500
one-way time [s]
500
1.0
a. subsurface model
two-way time [s]
-1500
0
b. time reversed focusing operator
0
1500
1
-1500
0
0
1500
1
2
-1500
0
0
0.5
-1500
0
0
1500
-1500
0
-1500
0
0
1500
0.5
1.0
f. synthesis for shot c)
1500
e. shot at x = 495
0.5
1.0
0
2
d. shot at x = 0
1500
one-way time [s]
-1500
0
1
2
c. shot at x = −495
one-way time [s]
0.5
1.0
g. synthesis for shot d)
-750
0
750
h. synthesis for shot e)
1500
0.5
➝
depth [m]
c=2000
1.0
i. CFP gather
Figure 5.9
Construction of a CFP gather for focusing in detection. Every common shot gather
contributes to one trace, positioned at the source position, in the CFP gather. Note the
contribution of the Fresnel zones in figure 5.9f, g and h to figure 5.9i. The focus point
response has been indicated with an arrow. Note also the relative simplicity of the CFP
gather.
Chapter 5: CFP technology
69
operator for the defined focus is shown in figure 5.9b. This operator is applied to all
common shot gathers available. Three different common shot gathers with source
positions at x = −495, x = 0 and x = 495 m are shown in figure 5.9c, d and e
respectively. Convolution along the time axis of the traces in the shot gathers with
the traces in the synthesis operator gives the intermediate synthesis results shown
in figure 5.9f, g and h. Note that in these intermediate synthesis results the bow-tie
of the syncline interface is still present. Summation over all the traces in the intermediate synthesis result defines one trace of the CFP gather. The most important
contribution in the integrated result is determined by the Fresnel zone related to
the focus point. If the focusing operator is correct then the operator time at the
source position is identical with the time of the event present in the CFP trace (for
a detailed discussion about erroneous operators see chapter 6). The summed trace
is placed in the CFP gather at the position of the source. By carrying out the convolution and integration along the traces in the gather for all shot gathers available
the CFP gather for focusing in detection is constructed.
The events which are present in the shot record are also present in the intermediate
synthesis result in figure 5.9f, g and h. In figure 5.9f four events are observed. The
top event originates from the first reflector and can be regarded as a ’non-causal’
event with respect to the focus. This event is ’non-causal’ because the focus is placed
below the first reflector, or said differently, the ’non-causal’ event appears before the
times of the used synthesis operator. The event with the triplication in it originates
from the syncline boundary, the weak S-shaped event originates from the diffraction
point and the last event originates from the deepest boundary. In the CFP gather
shown in figure 5.9i the reflection from the syncline (indicated by an arrow) and
the deeper boundary are clearly visible. The response of the first reflector is moved
outside the time window. Note that the bow-tie event in the shot gather (figure 5.9d)
is focused in the CFP gather (figure 5.9i) to a single event which is much simpler
to interpret. So due to the construction of the CFP gather the shot records are
transformed to a domain which is focused on only one grid-point in the subsurface
which can be less complicated.
In figure 5.10 another focus point is chosen in the syncline model. Now the contributions of the same shot records (c,d,e) in the CFP gather are different from figure
5.9. Only the deep reflector gives a contribution to the CFP response. Note that
before the CFP response a ’non-causal’ event is observed, which originates from the
syncline interface.
The event, indicated with an arrow, in the CFP gathers of figure 5.9i and figure
5.10i corresponding to the synthesis operator, called the focus point response, is
exactly equal in traveltime with the synthesis operator. This means that the used
synthesis operator contains no errors. If the source and receivers are defined at the
same positions in space then the same focusing operator can be used for the second
focusing step which operates on the CFP gather. In section (5.3) this second focusing
70
5.2 First focusing step
-1500
0
-750
lateral position [m]
0
750
1500
-1500
0
-750
lateral position [m]
0
750
1500
c=2200
1000
c=2700
●
c=2400
c=3000
1500
one-way time [s]
500
1.0
a. subsurface model
two-way time [s]
-1500
0
b. time reversed focusing operator
0
1500
1
-1500
0
0
1500
-1500
0
1
2
-1500
0
0
0.5
-1500
0
0
e. shot at x = 495
1500
0.5
1.0
one-way time [s]
-1500
0
-1500
0
0
1500
0.5
1.0
f. synthesis for shot c)
1500
2
d. shot at x = 0
1500
0
1
2
c. shot at x = −495
one-way time [s]
0.5
1.0
g. synthesis for shot d)
-750
0
h. synthesis for shot e)
750
1500
0.5
➝
depth [m]
c=2000
1.0
i) CFP gather
Figure 5.10
Construction of a CFP gather for focusing in detection. Every common shot gather
contributes to one trace, positioned at the source position, in the CFP gather. Note the
contribution of the Fresnel zones in figure 5.10f, g and h to figure 5.10i. The focus point
response has been indicated with an arrow.
Chapter 5: CFP technology
71
step is discussed in more detail. If the synthesis operator is in traveltime not equal
with the event of interest in the CFP gather the synthesis operator needs to be
updated, a process which is discussed in chapter 6. Note that by updating synthesis
operators (and not the initial macro model) a correct one-way time image of the
subsurface can be obtained without updating the initial macro model. To convert
this one-way time image to a structural model (a depth image) a macro model is
needed. This macro model can be estimated by using the collection of synthesis
operators and the obtained one-way time image. This tomographic inversion is
subject of another PhD project carried out in the Center for Technical Geoscience
by Rob Hegge (Hegge and Fokkema, 1996).
5.2.4
Focusing in emission and focusing in detection
In all previous shown CFP gathers, focusing in detection was carried out in the first
focusing step. The focusing operator was defined at the receiver positions and the
traces in the CFP gather were defined at the source positions. However, it is also
possible to do focusing in emission in the first focusing step. The focusing operator
must then be defined at the source positions and the resulting CFP gather will have
its traces positioned at the receiver positions.
To show clearly the differences between focusing in emission and focusing in detection
a flat reflector at 800 m depth with a velocity above the reflector of 2000 m/s is
chosen. The shot and receivers are moving above the reflector from left to right, the
first receiver is positioned at the source position and the last receiver is positioned
3000 m to the right of the first receiver, the distance between the receivers is 30
m. The shots are positioned at 30 m from each other in figure 5.11a,b and at 60 m
in figure 5.11c,d. Comparing figure 5.11a with 5.11b it is observed that due to the
marine type of acquisition geometry the focus is illuminated differently for focusing
in detection and focusing in emission.
In figure 5.11c the source spacing is 60 m and the receiver spacing is 30 m. Focusing
in detection gives a CFP gather sampled with traces 60 m (the shot distance) from
each other. Every trace is built from a shot record with the sampling distance of the
receivers. In the result for focusing in emission as shown in figure 5.11d the CFP
trace spacing is 30 m (the receiver distance). Every trace in this CFP gather is built
from a common receiver gather which is sampled with the source distance. Due to
the ’coarse’ source sampling the Fresnel zone is not properly sampled, which gives
the alias artifacts present in the top right hand-side of figure 5.11d. Note that these
artifacts are the same as the interference effects shown in figure 5.6.
The results of focusing in detection and focusing in emission can be combined into
one gather as shown in figure 5.11e and f. These combined CFP gathers are very
useful in the updating of erroneous focusing operators and the estimation of a macro
model (see chapter 6).
72
5.2 First focusing step
one-way time [s]
-1500
0
-750
lateral position [m]
0
750
1500
0.5
-1500
0
1500
1.0
a. focusing in detection ∆xs = 30 m
-1500
0
one-way time [s]
lateral position [m]
0
750
0.5
1.0
-750
0
750
1500
0.5
b. focusing in emission ∆xs = 30 m
-1500
0
-750
0
750
1500
0.5
1.0
1.0
c. focusing in detection ∆xs = 60 m
-1500
0
one-way time [s]
-750
-750
0
750
d. focusing in emission ∆xs = 60 m
1500
0.5
-1500
0
-750
0
750
1500
0.5
1.0
1.0
e. detection and emission ∆xs = 30 m
Figure 5.11
f. detection and emission ∆xs = 60 m
Focusing in detection and focusing in emission for different source sampling intervals.
Note the difference in illumination and the distance between the traces in the CFP
gather.
A simplified computation scheme for the first focusing step is shown in figure 5.12
where the difference between focusing in emission and focusing in detection is made
explicitly. The focusing operator defines the traveltimes from the focus to the source
and receiver positions which belongs to the input trace. The focusing blocks use this
traveltime to weight the input trace. The weighting of the input trace consist in
general of a time shift, a 45 degrees phase rotation and an amplitude scaling. The
weighted input trace is positioned into the CFP gather at the source and/or receiver
position. Other input traces which have their source or receiver position in common
with the previous input trace(s) will be added to the result.
In conclusion; focusing in emission synthesizes areal sources that aim at remote
illumination of single subsurface grid-points; the involved synthesis process trans-
Chapter 5: CFP technology
73
input trace
II
I
traveltime
to
receiver
traveltime
to
source
focusing
in
detection
focusing
in
emission
positioning
in
CFP gather
Figure 5.12
Block diagram for the traveltimes in the first focusing step. Every input trace can be
used for both focusing in detection (I) and emission (II). The weighted input trace is
positioned in the CFP gather.
forms field records into CFP gathers. Hence, each CFP gather defines an areal
shot record that represents the response of a focusing source. A single event in the
CFP gather may be interpreted as the response of a source at the defined grid-point
measured at the surface. Focusing the CFP gather gives a double focusing result.
A single event in the double focusing result may be interpreted as the output of a
simulated physical experiment; the response of a focusing source being measured by
a focusing detector with the same focus point (confocal) or with a different focus
point (bifocal). How the second focusing step is calculated will be explained in the
next section.
5.3
Second focusing step
The response of the focusing operator in the CFP gather contains information which
belongs to the same focus point. In the second focusing step this information is added
together and positioned in the one-way image gather. For this operation the focusing
operator which was already used in the first focusing step can be used again. Two
phases can be distinguished in this second focusing step; 1) move-out correction
with the focusing operator and 2) addition and positioning in the one-way image
gather. The first phase is obtained by a 1-dimensional convolution (in time) between
the CFP gather and its operator, the second phase is obtained by summation over
the traces in the CFP gather. Note that the combination of phase 1 and phase 2
74
5.3 Second focusing step
-750
lateral position [m]
0
750
1500
-1500
0
one-way time [s]
one-way time [s]
-1500
0
0.5
1.0
lateral position [m]
0
750
1500
0.5
1.0
a. focusing operator
-750
0
b. CFP gather
750
1500
one-way time [s]
-1500
-0.5
one-way time [s]
-750
0
0.5
-1500
-0.5
-750
0
750
1500
0
0.5
c. move-out corrected CFP gather
Figure 5.13
d. 2-dimensional cross-correlation (a**b)
The two phases in the second focusing step consists of a move-out correction (c) on
the CFP gather (b) using the focusing operator (a), followed by a summation over all
traces. In (d) the 2-dimensional cross-correlation between the CFP gather and its time
reversed operator is shown, the trace at x = 0 represents the image trace.
can be considered as the spatial zero-lag trace of a 2-dimensional cross-correlation
(in time and space) between the CFP gather and its time reversed operator. In
chapter 6 it will be shown that the 1-dimensional convolution result is important
for velocity analysis and operator updating, in section (5.4) it will be shown that
the 2-dimensional cross-correlation is important in AVO analysis. In this section the
imaging aspects of the second focusing step will be explained.
Figure 5.13 shows the second focusing procedure for the same focus point as defined
in figure 5.9. The time reversed focusing operator (a) time convolved with the CFP
gather (b) gives the move-out corrected CFP gather shown in figure 5.13c. For
a correct focusing operator all events of the focus point response in the move-out
corrected CFP gather occur at zero time. For the reflector at which the focus point
is positioned, this is indeed observed in figure 5.13c. The event occuring after zero
time is from a reflector which is positioned deeper than the focus point. From the
analysis of the time difference between the zero-time line and the response of the
deeper reflector a new operator can be estimated which should be correct for the
deeper reflector. An alternatively approach, to focus the deeper reflector, is that
given the time shifts between the event and the zero-time line the initial macro
model can be updated (differential time shift (DTS) analysis) and a new operator
can be calculated based on this updated model.
Chapter 5: CFP technology
75
The time reversed focusing operator (a) 2-dimensional cross-correlated with the CFP
gather (b) gives the so called reflectivity function (see section (5.4)) shown in figure
5.13d. The trace at x = 0 is the addition of all the traces in figure 5.13c and
represents, around zero time, the information of the focus point used in the imaging.
This information is positioned in the one-way image gather at a position defined
by the one-way image ray. Therefore the one-way image ray involved in the double
focusing technique must be defined first.
5.3.1
One-way image ray
In migration techniques there are several rays defined which play an important role
in time and depth migration (Hubral, 1977; Parkes and Hatton, 1987). Parkes
and Hatton (1987) define the image ray as ”the ray associated with the minimum
traveltime from a subsurface point to the surface”. The normal ray is associated with
the minimum traveltime from a coincident source/receiver pair to any particular
interface. By definition the normal ray is perpendicular to the target interface and
the image ray is perpendicular to the surface. Depth migration of data has the effect
of moving points along their image ray to their correct position. Time migration of
data has the effect of moving points laterally to their minimum time positions (the
normal ray), rather than their ’true’ time positions.
By the CFP technology a new ray is introduced: the one-way image ray. The oneway image ray is associated with the traveltime of a focus point in the subsurface to a
receiver at the same lateral position as the focus point at the surface. In figure (5.14)
the one-way image ray is displayed together with the image ray and the normal ray
(in order to observe a better difference between the different ray paths the velocities
of the syncline model are modified). Note that if the image ray is used for imaging
depth [m]
-1500
0
-750
ray y
ray ay l ra
ge e-w rma
a
im on n0o
750
1500
500
1000
1500
Figure 5.14
Normal, image and one-way image ray for a focus point in the subsurface in the syncline
model. The one-way image time is given by the time of the zero CFP offset trace in the
synthesis operator. Note that in general the one-way image ray is bended.
76
5.3 Second focusing step
a lateral shift of the data points is needed to position the reflector at its correct
position. With the definition of the one-way image ray the result of the double
focusing procedure for one focus point is positioned in the one-way image gather at
the lateral position of the focus point at the one-way image time. Therefore in the
CFP imaging procedure the one-way image ray is, by definition, positioned at the
correct lateral position and there is no need to shift the data points in the conversion
to a depth image.
The one-way image ray defines where the result of the double focusing procedure
must be positioned in the one-way image gather. The positioning of the double
focusing result is explained in detail in the next subsection.
5.3.2
Positioning of the double focusing result
The focusing operator is defined in one-way traveltime. The synthesis process converts the two-way time of the data to a mixed version of one-way and two-way
traveltimes. The one-way part of the CFP gather is located at the operator times.
The events in the CFP gather above and below the operator times are associated
with a two-way traveltime with respect to the defined focus point. In figure 5.15 this
is explained with a 1 dimensional model with 3-reflectors (a). The velocity in the
1-dimensional medium is chosen constant at c = 2000 m/s and at 400, 800 and 1200
m(= 0.4 seconds difference in two-way traveltime), the reflectors in the model are
defined by a density contrast. The CFP gather for the first ➊ and third ➌ reflector
are shown in figure 5.15d and f and the focusing operators are shown in c and e
respectively. Note that the focusing operators are time coincident with the related
response in the CFP gather, indicating that the correct focusing operators are used.
In the CFP gather for the third reflector (f) the response of the first reflector ➊
is shifted to negative times and is located at the bottom of the picture. Looking
at the traveltimes for the trace at x = 0 in the zoomed pictures (g) and (h) it is
observed that the time between the position where the focus point is defined and the
other layers is the two-way interval traveltime between the layer and the point (0.4
seconds for x = 0). The time of the focus point response is the one-way traveltime
of the focus point to the surface. This mixture of one-way and two-way traveltimes
must be taken into account for a proper positioning of the events in-between two
focus points defined at the same lateral position but at different one-way image time
levels.
Suppose that there are only two focus points defined; one at the first and one at
the third reflector as shown in figure 5.15a and that the aim is to image the middle
reflector ➋. Using the one-way image ray and the second focusing step the imaging
procedure can be carried out for the two focus points. Figure 5.16 shows the moveout corrected CFP gather for the first reflector (a) and for the third reflector (b).
According to the one-way image ray the first reflector must be imaged at the one-
Chapter 5: CFP technology
77
way time of 0.2 seconds, the second reflector at 0.4 seconds and the third reflector
lateral position [m]
0
750
1500
➊
➋
1000
-1500
0
two-way time [s]
-750
500
➌
1
a. subsurface model
one-way time [s]
-1500
0
lateral position [m]
0
750
1500
➊
➋
➌
-750
0
b. shot record at x = 0 m
750
1500
1
-1500
0
1
2
-750
0
750
1500
➊
➋
➌
2
c. focusing operator for ➊
-1500
0
one-way time [s]
-750
2
1500
-750
0
d. CFP gather response for (c)
750
1500
-1500
0
-750
0
750
1500
➊
1
➋
1
➌
2
2
e. focusing operator for ➌
-1500
0
one-way time [s]
depth [m]
-1500
0
0.4
0.8
-750
750
1500
-1500
0
0.4
➊
➋
0.8
g. zoom of (d)
Figure 5.15
0
f. CFP gather response for (e)
-750
0
750
1500
➋
➌
h. zoom of (f)
Traveltimes in the CFP gather. The time between the focus point response and the
other reflectors is the two-way interval traveltime. This is observed in (g) where the
800
time between ➊ and ➋ at x = 0 is a two-way traveltime (= 2d
= 2000
= 0.4s).
c
78
5.3 Second focusing step
one-way time [s]
-1500
-1.0
-750
lateral position [m]
0
750
1500
-1500
-1.0
-0.5
-0.5
0
0
0.5
0.5
1.0
lateral position [m]
0
750
1500
1.0
a. move-out correction for ➊
-1500
0
one-way time [s]
-750
-750
0
b. move-out correction for ➌
750
1500
0.5
-1500
0
-750
0
750
1500
0.5
1.0
1.0
c. conversion to one-way time of (a)
one-way time [s]
-1500
0
d. conversion to one-way time of (b)
lateral position [m]
-750
0
750
1500
0.5
1.0
e. CFP image gather
Figure 5.16
Second focusing step and positioning at the one-way image time. The CFP image gather
is obtained by combining the results of (c) and (d) and using interpolated operators to
correct for the move-out in between the defined focus points.
at 0.6 seconds. Note that in figure 5.16a the time between the response of the first
and the second reflector is 0.4 seconds. To image the second reflector at the correct
one-way time (0.4 s), by using the CFP gather with a focus point defined at the first
reflector, the two-way time between the second reflector and the focus point should
be converted to one-way time. The result of this conversion is positioned at the
one-way image time of the focus point and is shown in 5.16c for the first reflector
and in 5.16d for the third reflector.
Note that at the one-way image time of the focusing operator (0.2 s) in figure 5.16c
the move-out corrected response for the first reflector is flat and the response of the
second reflector (around 0.4 s) is not flat. The non-flatness of the second reflector
is due to the fact that the wrong operator (defined only for the first reflector) is
Chapter 5: CFP technology
-750
lateral position [m]
0
750
1500
500
1000
∇∇∇∇∇∇∇∇∇∇∇∇∇∇∇∇∇∇∇
1500
-1500
0
one-way time [s]
depth [m]
-1500
0
79
-750
lateral position [m]
0
750
1500
0.3
0.6
a. syncline model
Figure 5.17
b. CFP-stack at z = 1200 m
The CFP-stack gives a local image around the defined focus points. Note that all the
other events present in the syncline model, like the syncline interface and the wedge,
are imaged at the correct one-way image time as well.
used for the move-out correction in the second focusing step. Therefore one should
use in the second focusing step, for every different image-time a different operator.
These intermediate operators can be obtained by interpolating the operators used in
the first focusing step. How sparse the operators can be chosen in the first focusing
step is investigated in section (5.4) where the resolution and focusing beam of the
operator is analyzed.
The results of the second focusing step combined into one gather is shown in figure
5.16e and is called the CFP image gather. Interpolated operators are used to correct
for the move-out in between the defined focus points. There is a time window used to
select the contribution from the CFP gather with a focus point on the first reflector
which is determined by the difference in one-way image time between the focusing
operator at the third reflector. This time window around the focus point response is
chosen such that the contribution of the CFP gathers below and above just overlap.
Figure 5.16e is obtained by combining the response of the first and third reflector.
The constructed CFP image gather is of special interest and will be discussed in the
next section.
Another useful gather is the CFP stack. The CFP-stack is defined as a move-out
corrected CFP-gather converted to one-way time followed by a summation over all
the traces in the gather. In the CFP-stack the event at the image time of the defined
focus point is equal to the double focusing result. Around the one-way image time
the CFP-stack will give a good representation of the area surrounding the focusing
point. The CFP-stack can be used to get quick an idea of the main reflectors in the
data, but it can also be used to investigate a particular reflector or depth level.
For example by calculating CFP gathers with focus points following, in the lateral
direction, a horizon or a reflector, followed by the CFP stacking procedure builds up
an image for the horizon or boundary of interest. Figure 5.17b shows a CFP-stack
for focus points defined at the deep flat reflector in the syncline model. Note that
80
5.3 Second focusing step
-1500
0
-750
lateral position [m]
0
750
1500
two-way time [s]
500
∇
1000
∇
∇
-750
lateral position [m]
0
750
1500
➊
➋
➌
➍
1
2
1500
a. subsurface model
b. CMP gather
lateral position [m]
-750
0
750
-1500
0
one-way time [s]
depth [m]
∇
-1500
0
1500
➊
➋
➌
➍
0.5
1.0
-1500
0
-750
0
750
1500
➊
➋
➌
➍
0.5
1.0
c. X-gather
Figure 5.18
d. CFP image gather
For the calculation of a CFP image gather at one lateral position the response of several
focus points (in depth) are combined in one gather. Note that if the correct operators
are used the events in the CFP image gather all line up horizontally.
all the other events present in the syncline model, like the syncline interface and
the wedge, are imaged at the correct one-way image time as well. This remarkable
property of the CFP method is discussed in section (6.4), but first the CFP image
gather and the CFP image must be discussed in more detail to be able to fully
appreciate this property of the CFP method.
5.3.3
CFP image-gather
The CFP gather gives the response from one focus point in the subsurface. The
most important information in the CFP gather is therefore concentrated around the
focus point response. The CFP image gather is constructed by selecting samples
from the CFP gather around a time window, defined by the focusing operator, and
placing these selected time samples into one gather. This procedure is repeated
for different CFP gathers constructed with focusing operators defined at the same
lateral position, but at different depths (or one-way times). In the CFP image
gather the samples from the CFP gather are converted to one-way time with respect
to the one-way image time of the focusing operator. For the move-out correction in
the CFP image gather at every one-way time sample an operator is needed. These
operators are obtained by a linear interpolation (in offset or in ray-parameter p) of
the operators used in the first focusing step. With these operators the CFP image
gather can be move-out corrected and stacked to get the image trace. This move-
Chapter 5: CFP technology
81
out corrected multi focus gather represents the CFP image gather in one-way image
time. If the move-out correction is not applied the gather is called an X-gather (X
from exploding).
In figure 5.18d a CFP image gather, positioned at x = 500 m, is shown for the
syncline model. The image gather is constructed from 4 CFP gathers with their
focus points defined at the boundaries of the model (a). Event ➊ originates from
the flat top reflector, ➋ from the right flank of the syncline (note the finite aperture
artifact due to the dip at the flank), ➌ originates from the deep flat reflector (note
the negative reflection coefficient) and ➍ from the wedge in the right part of the
model. The X-gather (c) can also be used as an alternative for the CMP gather
which is shown for comparison in 5.18b.
Building the CFP image gather, by making use of traveltime operators, introduces
a stretch of the wavelet due to the convergence of the operator times at the higher
offsets. This stretched part at the higher offsets can be removed from the CFP image
gather by setting a stretch parameter. In figure 5.18d the stretched wavelet at the
higher offsets is clearly visible.
coherency measurement
In the CFP image gather all events should align at the same one-way image time, if
an event is not aligned this indicates that an erroneous focusing operator was used
or that the response of the focus point was not measured. Any misalignment with
respect to a reference trace can be calculated and used as a quality measurement
of the image. Therefore the following coherency measurement is introduced (after
Neidell and Taner, 1971)
Cj = s
j+N/2
P
Ak Bk
k=j−N/2
j+N/2
P
k=j−N/2
A2k
j+N/2
P
k=j−N/2
,
(5.7)
Bk2
where Cj is the coherency measurement at time sample j, A a trace of the CFP
image gather and B the reference trace. The ideal choice for B is the image trace
itself, an alternative would be the image trace which is built up during the imaging.
5.3.4
One-way time image
The CFP-stack, which was explained in the previous subsection, represents a local
image around the defined focus points. By defining focus points not only at one
boundary but in the whole subsurface a complete one-way time image can be built
up. As mentioned before it is not necessary to define in the first focusing step a
focus point at every time position for all lateral positions. Using the resolution
in the focusing beam an optimum focus point distribution can be determined (see
82
5.3 Second focusing step
-750
lateral position [m]
0
750
1500
➊
1
➋
➌
➎
-1500
0
one-way time [s]
two-way time [s]
-1500
0
2
lateral position [m]
0
750
1500
0.5
1.0
a. shot record for a source at x = −510 m
-750
0
750
➋
➌
➍
0.5
b. synthesis operator at x = 500, z = 1200 m
-1500
0
-750
0
750
1500
0.3
➝
1.0
1500
one-way time [s]
➝
-1500
0
one-way time [s]
-750
0.6
c. CFP gather
Figure 5.19
d. one-way time image
The contribution from a shot record (a) in the CFP image (d) depends on the used
focusing operator (b). Note that in this example only a small part of the shot record
contributes to the image position (indicated with an arrow).
section (5.4)). For example a good choice would be a focusing operator positioned
at every macro boundary in the model.
The imaging procedure is explained in figure 5.19 by using the data of the syncline
model. For the imaging procedure a focus point is chosen at every 15 m at the
three main boundaries in the model and at the diffraction point. The shot record
shown in figure 5.19a has a contribution in the CFP gather (constructed with the
focusing operator shown in figure 5.19b) at its source position x = −510, which is
indicated by the arrow in 5.19c. Note that event ➎ in the shot record originates
from the diffraction point. The focusing operator (b) has its focus point defined
at x = 500, z = 1200. Performing a second focusing step on the CFP gather with
the same operator of figure 5.19b and selecting the samples around the operator
times gives the contribution indicated by the arrow in the one-way time image of
figure 5.19d. The lateral position in the image is by definition equal to the lateral
position of the focusing operator. The imaging for the samples in between two
focusing operators at the same lateral position is done with interpolated focusing
operators (the second focusing step). Along the times given by these second focusing
operators the samples are stacked and positioned at the one-way image time. The
time window, which determines which samples are included from the CFP gather
in the image, is determined by the time difference between the two operators. Note
that due to the fixed spread acquisition geometry the edges of the model are sparsely
Chapter 5: CFP technology
one-way time [s]
-1500
0
-750
lateral position [m]
0
750
1500
0.3
-1500
0
-750
lateral position [m]
0
750
1500
0.3
0.6
0.6
a. homogeneous model and 1 layer
-750
0
750
b. homogeneous model and 3 layers
1500
depth [m]
-1500
0
one-way time [s]
83
0.3
0.6
-1500
0
-750
0
750
1500
500
1000
1500
c. correct model
Figure 5.20
d. depth image with correct model
Imaging results for the double focus procedure using synthesis operators based on a
homogeneous model; with only one level of focus points (a), with three levels of focus
points (b) and the correct model (c). Result (d) shows the pre-stack depth migration
result using the correct macro model. Note that result (a) means stacking of CFP gathers.
illuminated in the image (d). Note also that the imaging condition used in the CFP
method states that a reflector exists within the earth when the wavefront curvature
of the downward continued wavefield is the same as the curvature of the downward
incident wavefield for all offsets.
To show the robustness and sensitivity of the imaging procedure the same imaging
procedure is carried out with synthesis operators which are based on a erroneous
initial macro model. For the erroneous macro model a homogeneous model is chosen
with the velocity c = 2200 m/s. Figure 5.20a shows the imaging result by using the
homogeneous model with only one layer defined at z = 800. The focusing operators
are not updated during the imaging procedure. Figure 5.20b shows the imaging result with three depth levels at z = 400, z = 800 and z = 1200 m in the homogeneous
model where the focusing operators are defined. For a good comparison the imaging
result using the correct model is shown in figure 5.20c and the result obtained with
conventional prestack depth migration using the correct model is shown in figure
5.20d. Note that the one-way time image obtained with the homogeneous model
gives already a good indication of the reflectors present in the model. The images
obtained with the homogeneous macro model show also that the reflectors are all
imaged at the same one-way image time, there is only a difference in amplitude.
84
5.3 Second focusing step
two-way time [s]
-1500
-750
lateral position [m]
0
750
1500
-1500
-0.4
-0.4
-0.2
-0.2
0
0
0.2
lateral position [m]
0
750
1500
0.2
a. image built with ∆x = 30 m.
-1500
two-way time [s]
-750
-750
0
750
b. image built with ∆x = 90 m.
1500
-1500
-0.4
-0.4
-0.2
-0.2
0
0
0.2
-750
0
750
1500
0.2
c. image built with ∆x = 150 m.
Figure 5.21
d. image built with ∆x = 210 m.
CFP imaging by a summation of bifocal images for different lateral spacings; the image
of the third boundary (t = 0) is aimed for only.
This same effect was already observed in the discussion about the CFP-stack and
will be explained in section (6.4). Comparing the depth migration with the one-way
time image shows that the conversion from the one-way time image to the depth
image consists only of a (non-linear) stretching of the time axis, a lateral shift is not
needed. The one-way time image can be calculated with any set of focusing operators, which makes it a suitable method for analysis in a special regions of interest.
The CFP imaging method can also be implemented in an efficient way by making
use of a processing method which is done entirely in the time domain (see appendix
C).
The one-way time image can also be built up in another more efficient, but less
accurate, way. After the construction of the CFP-gather, it can be cross-correlated
(2-dimensional) with its time reversed synthesis operator. The cross-correlation can
be interpreted as a double focusing result where it is assumed that the reflector is
locally flat. The result of this cross-correlation is placed in the image gather at the
lateral position of the focusing operator. In the imaging procedure discussed above
only the spatial zero-lag trace (confocal result) of the 2D cross-correlation was placed
in the imaging result, now every trace (bifocal result) is placed in the image. In this
way the spatial width of the bifocal image can be used in an optimal way and it
is not necessary to use a laterally dense sampled set of focus points. The result of
this imaging procedure is shown in figure 5.21 for different spatial sampling rates
between the focusing positions on the deep flat reflector. Note that the one-way time
Chapter 5: CFP technology
85
re-sampling is not carried out in the examples shown in figure 5.21, so the data is still
in two-way time. Note also that in figure 5.21 the samples around t = 0 represent
the result which can be used in the one-way time image, the synclinal event which
occurs at negative times can not be used directly in the one-way time image. For a
complete construction of the one-way time image with the cross-correlation method,
at every macro boundary focus points must be chosen and combined to construct
the image.
The images shown in figure 5.21 can be interpreted as a plane wave (normal incidence
on the reflector) response measured at the deep flat reflector in the syncline model. In
this interpretation both the sources and receivers are positioned at the deep reflector.
Using this interpretation one can migrate the constructed image with an areal shot
record migration algorithm (Rietveld, 1995). The obtained migration result will not
be as good as a ’conventional’ areal shot record migration because the assumption
is made that the operators can be shifted laterally, which means that the medium
is 1-dimensional.
5.4
Resolution and amplitude analysis
Thus far no attention has been paid on the resolution of the focusing method. In
the discussed imaging procedure it was implicitly assumed that the focusing area
around the focus point has an extent in the vertical (time) direction. By using this
assumption it was argued that it is not necessary to place at every time sample a
focusing operator in the first focusing step. The resolution analysis in this section
will show that this assumption is indeed valid. Further it will be shown that besides
the imaging properties, the focusing method also provides a method of determining
the reflection coefficient.
5.4.1
Resolution and focusing beams
Modeling the energy for a propagating wavefront of the synthesis operator through
the subsurface model gives an indication how the areas in the subsurface are illuminated by this synthesis operator. From this so called focusing beam it is also possible
to determine which aperture at the surface is most important for the illumination.
With these illumination areas in the subsurface the distribution of the focus points
in the subsurface can be determined in order to obtain an efficient and optimum
illumination procedure of the subsurface. In figure 5.22 three focusing beams are
shown for the syncline model introduced earlier. The beams are constructed by performing an inverse recursive depth extrapolation of the focusing operators through
the model and calculating at every depth level the energy of the wavefield as function of the lateral position. Note that for the construction of the beams only the
synthesis operators and a macro model are needed.
86
5.4 Resolution and amplitude analysis
depth [m]
-1500
0
-750
lateral position [m]
0
750
1500
-1500
0
1500
750
➊
1000
➌
1500
1500
b. focusing beam for position ➊
a. subsurface model
-1500
0
depth [m]
lateral position [m]
0
750
500
➋
-750
0
750
1500
750
-1500
0
-750
0
750
1500
750
1500
1500
c. focusing beam for position ➋
-1500
0
depth [m]
-750
-750
0
750
750
d. focusing beam for position ➌
1500
-1500
0
-750
0
750
1500
750
1500
1500
e. snapshots for position ➋
Figure 5.22
f. snapshots for position ➌
Focusing beams through the subsurface model given in a). Note the tube like shape of
the different focusing areas.
From the focusing beams in figure 5.22 it is observed that around the actual focus
point most of the energy is focused in a tube like shaped area. The black center of the
tube has a lateral extent which is smaller than the vertical extent. The shape of the
focusing area is directly related to the resolution at the focusing point. The lateral
and vertical resolution of a focus point are determined by the acquisition geometry
and the subsurface model. By using this information combined with the information
from the focusing beams it is possible to define at which vertical sampling density
the focus points have to be chosen to illuminate the subsurface properly.
In figure 5.23 a zoomed version of the focusing area in figure 5.22c is shown. Figure
5.23a shows that the focusing area is not limited to the defined focus point but has
an extension in the horizontal and vertical direction. In the contour plot, plotted
with equidistant contour values, it is observed that the inner contour lines are close
Chapter 5: CFP technology
-750
-500
-750
-250
1000
1000
1200
1200
1400
1400
a. grayscale plot
Figure 5.23
-500
87
-250
b. contour plot
c. 3D plot
Details of the focusing beam for focusing position ➋ in figure 5.22c. Note that the
focusing area is not limited to the defined focus point at x = −500, z = 1200 m.
together. In the 3D plot this is observed as a broad peak at the focus point position.
For this example a good illumination in the horizontal direction can be obtained by
placing at every 50 m a focus point and a good illumination in the vertical direction
can be obtained by placing a focus point at every 100 m.
In figure 5.24 different combinations of focusing operators are used to illuminate
the subsurface. The illumination shown in figure 5.24a is obtained by placing three
focus points at the same lateral position in the model (one at every boundary). This
illumination is not sufficient to illuminate the events in-between the boundaries.
Reducing the distance between the focus point gives the results shown in figure 5.24b
-1500
0
-750
0
750
1500
750
-1500
0
-750
0
750
1500
750
1500
-1500
0
-750
0
1500
750
-1500
0
-750
0
-1500
0
-750
0
1500
750
1500
1500
-1500
0
-750
0
0
750
1500
750
1500
f. ∆x = 30 m
750
1500
-1500
0
-750
0
750
1500
g. aperture = 2000 m
-750
1500
750
Figure 5.24
-1500
0
e. ∆x = 60 m
750
1500
750
1500
d. ∆x = 90 m
750
c. ∆z = 120 m
750
750
1500
0
1500
b. ∆z = 225 m
750
-750
750
1500
a. ∆z = 450 m
-1500
0
1500
h. aperture = 1000 m
i. aperture = 500 m
Different focus point distributions in lateral and vertical direction. Note that for a
complete illumination a sparse distribution of the focus points is sufficient.
88
5.4 Resolution and amplitude analysis
depth [m]
-1500
0
-750
lateral position [m]
0
750
1500
-1500
0
lateral position [m]
0
750
1500
500
750
➊
1000
➋
➌
1500
1500
b. focusing beam for position ➊
a. subsurface model.
-1500
0
depth [m]
-750
-750
0
750
750
1500
-1500
0
-750
0
750
1500
750
1500
1500
c. focusing beam for position ➋
Figure 5.25
d. focusing beam for position ➌
Focusing beams of CFP gathers extrapolated through the subsurface model given in a).
Note that there are more focusing areas than in figure 5.22.
and 5.24c. From these experiments it can be concluded that in the first focusing
process focus points distributed with a vertical distance of ±100 m are sufficient
to illuminate the events in-between the focus points correctly. This is of significant
importance for efficiency, particularly in 3D.
The horizontal distance between the two focus points must be chosen smaller than
the vertical distance to obtain a sufficient illumination. In figure 5.24d,e and f three
different horizontal focus point distributions are shown. The distance of ∆x = 60 m,
shown in figure 5.24e, is still not sufficient for a proper illumination. Another way
to influence the horizontal illumination is by limiting the operator aperture such the
focusing beams become broader in the lateral direction as shown in figure 5.24g,h
and i. By limiting the operator aperture (which means reducing the maximum
angle which can be focused) a broader focusing beam is obtained which gives a less
accurate image, but this image can be obtained by using less operators.
The introduced analysis with focusing beams are a helpful tool in determining the
focus point distribution in the subsurface. The focusing beams give more information
than ray-traces, because the resolution of the focusing energy is shown as well. The
same beam-focusing algorithm can also be used for monitoring the focusing of all
the events present in the constructed CFP gather. In figure 5.25 the focusing beams
are shown for the CFP gathers constructed with the operators used to construct
the beams of figure 5.22. At the defined focus point the focusing beam should
be the same as the focusing beam of the operators. Due to the presence of the
Chapter 5: CFP technology
89
angle dependent reflection coefficient in the CFP gather, figure 5.25b gives a slightly
different focusing pattern as in figure 5.22b. Beside a beam at the defined focus
point the other out of focus reflections are also visible. Below the main focus point
in figure 5.25b there are two focus areas visible which originate from the deeper
reflections. Note that these focus areas are not positioned on a boundary. In figure
5.25c (Figure 5.10i shows the construction of the CFP gather used in the calculation
of the beam in figure 5.25c) a strong focusing area is present above the defined focus
point. The strong ’non-causal’ event in the CFP gather above the traveltime of
the operator gives rise to the observed strong focusing area. This area cannot be
interpreted physically because it originates from a ’non-causal’ event present in the
CFP gather. If one wants to do analysis on beams calculated from CFP gathers,
the ’non-causal’ events in the CFP gather should be muted out. In figure 5.25d it
is observed that another focusing area is present at a deeper depth level due to the
presence of the deeper reflector.
5.4.2
Amplitude analysis
If the correct focusing operator is used the time-reversed focusing operator and the
focus point response have equal traveltimes. The only difference between the focus
point response and its time-reversed focusing operator is in amplitude. According
to equations (4.33) and (4.31), which are repeated below,
∗ −1
Fi− (zm , z0 ) = Ii− (zm ) W+ (zm , z0 ) D− (z0 )
(5.8)
Pi− (zm , z0 ) = Ii− (zm )R+ (zm )W+ (zm , z0 )D+ (z0 )S(z0 )
(5.9)
the difference in amplitude along the traveltime curve is given by the angle dependent
reflection property at the focus point Ii− (zm )R+ , the surface operators D± and the
source signature S(z0 ). Assuming that the source signature is the same for all
experiments S(z0 ) = IS and that there are omni-directional receivers at the surface
D− (z0 ) = I it is possible to derive a simple procedure which can extract the reflection
information from the CFP gathers and its focusing operator. Note that in equation
(5.9) the limited spatial bandwidth is represented by the (limited) size of the D
matrices. Before the amplitude analysis in a CFP gather is discussed in detail first
the amplitudes for an individual trace of a shot record, a focusing operator and a
CFP gather are considered.
The amplitudes present in the CFP gather can be interpreted by looking at the
construction of the CFP gather for the top flat layer in the syncline model shown
in figure 5.9a. The synthesis operator is defined for a focus point at the reflector at
x = 0 and z = 300 m. The stationary phase contribution in the CFP gather of the
shot record with a source position at x = −300 is given by the trace at x = 300. This
stationary phase trace is shown at the left hand side of figure 5.26. The amplitude
RS
, where S represents the amplitude of the wavelet, R
in this trace consists of √
2r
5.4 Resolution and amplitude analysis
0.8
Figure 5.26
-0.06 0
0
A≃
RS
√
2r
0.4
0.8
0.06
-0.02 0
0
A≃
S
√
r
0.4
0.02
CFP gather
0.4
0.02
shot gather
-0.02 0
0
operator
90
A≃
RS
√
r
0.8
Amplitudes of a trace from a shot gather, synthesis operator and CFP gather, where S
represents the amplitude of the wavelet, R(= 0.29) the reflection coefficient at 45◦ and
r the distance between the focus point and the receiver at the surface.
the reflection coefficient at 45◦ and r the distance between
the focus point and the
√
receiver at the surface (in the given example r = 2 ∗ 3002 ). The operator trace
at x = 300, the middle trace in figure 5.26, has an amplitude proportional to √1r .
The trace on the right hand side of figure 5.26 represents the contribution of the
shot gather in the CFP gather. The amplitude in the CFP gather is proportional
√ . Compared with the trace in the shot record there is a strong amplitude
to RS
r
effect due to the summation in the Fresnel zone. Dividing (in a least-squares sense)
the CFP gather by its synthesis operator, weighted with the wavelet (the second
trace in figure 5.26) gives the reflection coefficient for an angle of incidence of 45◦ .
This relationship between the CFP gather and the synthesis operator can be used to
determine the AVO behavior at the focus point. Before this procedure is explained
in detail first another aspect of the amplitudes is discussed.
From the foregoing discussion it can be concluded that it is possible to quantify
the angle-dependent reflection property of the focus point by computing for each
offset the amplitude ratio between the focus point response and the time-reversed
focusing operator. This is illustrated for a flat reflector in figure 5.27; the reflector
is positioned at 800 m depth and an acoustic contrast is used. The amplitude ratio
shown in figure figure 5.27d has been computed in the time domain in a least-squares
manner: for each offset the zero-lag temporal cross-correlation between the response
and the focusing operator is divided by the zero-lag temporal autocorrelation of the
operator. The critical reflection angle occurs at an offset of 1060 m, which is the
starting point for the estimated AVO curve to deviate from the theoretical curve.
AVP (amplitude versus ray-parameter) analysis is more suitable for angle-dependent
reflection purposes than AVO analysis. For AVP analysis the bifocal version of
CFP migration is required (Berkhout, 1996b). In CFP imaging the confocal result
(2-dimensional cross-correlation between the time-reversed operator and the CFP
gather with only a zero-shift in the spatial direction) of the second focusing step is
used only, for AVP analysis the bifocal result (2-dimensional cross-correlation with
Chapter 5: CFP technology
-1500
0
-750
lateral position [m]
0
750
1500
750
two-way time [s]
depth [m]
c = 2000 [m/s]
ρ = 2000 [kg/m3 ]
-1500
0
●
c = 2500 [m/s]
ρ = 1000 [kg/m3 ]
91
-750
lateral position [m]
0
750
1500
1
2
1500
a. flat reflector model
-750
0
750
1500
1.0
amplitude versus offset curve
theory →
0.8
amplitude
one-way time [s]
-1500
0
b. shot record with xs = 0
0.5
0.6
0.4
0.2
CFP →
0
-0.2
1.0
-0.4
c. CFP gather
Figure 5.27
-750
0
750
1500
1.0
-750
-0.2
1500
-375
0
375
750
0
0.2
a. CFP gather
0
10
ray parameter [ms/m]
20
30
b. 2D cross-correlation (bifocal image)
40
50
1.0
amplitude versus ray-parameter curve
theory →
0.8
amplitude
intercept time [s]
1000
Amplitude versus offset curves obtained by a least-squares estimation in the time domain. Note that the critical reflection angle occurs at an offset of 1060 [m].
0.5
-0.2
500
offset [m]
one-way time [s]
one-way time [s]
-1500
0
0
0
0.6
0.4
0.2
CFP →
0
-0.2
0.2
-0.4
c. τ − p transformation of b)
Figure 5.28
0
10
20
30
ray parameter [ms/m]
40
50
Amplitude versus ray-parameter curve obtained by a τ −p transformation of the bifocal
image. Note that the critical reflection angle occurs at an ray-parameter of 40 × 10−3
[s/m].
92
5.4 Resolution and amplitude analysis
shot
records
focusing in
detection
CFP gather
operator
adjustment
bifocal
Figure 5.29
confocal
focusing in
emission
focusing in
emission
angle-dependent
reflectivity
angle-averaged
reflectivity
Diagram for bifocal imaging and AVO analysis. After the first focusing step bifocal and
confocal imaging can be carried out resulting in different reflectivity
’all’ shifts in the spatial direction) is used. The bifocal result of the CFP gather is
shown in figure 5.28b, the trace at x = 0 is used in the confocal imaging procedure.
By performing a linear Radon (τ −p) transformation on the bifocal image, the bifocal
image is decomposed in local plane waves (ray-parameters) (c). The amplitude at
τ = 0 (d) of every ray-parameter represents the amplitude of the (band limited)
reflection operator at the focus point. The critical reflection angle occurs at the
ray-parameter value of 40 × 10−3 s/m, which is the starting point for the estimated
AVP curve to deviate from the theoretical curve.
The flow scheme of the bifocal imaging procedure is shown in figure 5.29. After
focusing in detection and optionally updating of the focusing operator the second
focusing step can be carried out. The integration of velocity analysis, imaging and
AVO analysis can all be combined within the same double focusing process. The
analysis and processing modules are included between the two focusing steps. This
Chapter 5: CFP technology
93
does not only allow a better control on the structural solution and the reservoir
characterization problem, its also offers new solutions to the well known surfacerelated problems (Berkhout and Verschuur, 1996b).
In the foregoing examples the amplitude analysis has only be shown for a single flat
reflector. AVO analysis of more complicated examples is beyond the scope of this
thesis and subject of another PhD project carried out by Aart-Jan van Wijngaarden.
5.5
3-Dimensional CFP gathers
As might be expected the CFP technology can also be used to process three dimensional seismic data. However, due to the irregular sampling in the cross-line
direction special care should be taken in the first focusing step to obtain an appropriate summation in the Fresnel zone. If the Fresnel zone is positioned outside the
seismic aperture or if the sampling is inaccurate the contribution in the CFP trace
originating from the Fresnel zone will get distorted. To investigate the influence
of irregular and sparse sampling first some 2-dimensional common offset contributions of the CFP gather are discussed. After this discussion the construction of a
3-dimensional CFP gather is much easier to understand. At the end of this section a method is proposed which can regularize CFP gathers in the coarse sampled
(cross-line) direction.
5.5.1
Common offset contributions in 2-dimensional synthesis
Suppose that a CFP gather is constructed from shot records which contain only
one trace. Then the CFP gather will consist of this single trace convolved with
the appropriate trace of the focusing operator. In figure 5.30 this is shown for two
offsets for the model used before; a single flat reflector at 800 m depth and sources
moving from −1500 m to 1500 m with an interval of 15 m. The focusing operator
is defined for a focus point positioned at the reflector at x = 0. In figure 5.30b
the contribution of the zero offset trace is shown and in (d) the contribution of
a trace with an offset of 1005 m is shown. Note that the common-offset section
of a flat reflector gives a flat response. Convolution with the focusing operator and
positioning at the source position gives the results shown in figure 5.30. The focusing
operator can be recognized easily in the result of the zero-offset contribution. The
traveltimes of the time-reversed focusing operator are represented by the spiky event.
The contribution from the receiver with an offset of 1005 m is tangent with the time
m, which originates
reversed focusing operator at the lateral position x = −1005
2
from a source position at 1005
m.
2
By taking more offsets into account the CFP gather will be built up completely. It
is interesting to investigate how many offsets must be taken into account to build up
the CFP gather completely. The best way to compare the different results, obtained
94
5.5 3-Dimensional CFP gathers
-1500
0
●
one-way time [s]
∇
∗→
-750
lateral position [m]
0
750
1500
0.5
1.0
a. zero-offset
-1500
0
∇
●
one-way time [s]
∗→
b. zero-offset contribution
-750
lateral position [m]
0
750
1500
0.5
1.0
c. 1005 [m] offset
Figure 5.30
d. 1005-offset contribution
Common offset contribution to the CFP gather. The traveltimes of the time-reversed
focusing operator are represented by the spiky event. Note that the focusing operator
can be recognized in the contribution of the single offset.
with different number of offsets, is with the aid of the one-way image trace. The
number of offsets which are taken into account are varied by changing the detector
spacing. In figure 5.31 several source and receiver sampling rates are used to build up
the image trace. Beside the detector spacing the source spacing can also be varied.
The acquisition spread is moving together with the source distance, like in a marine
acquisition geometry. The sources are ranging from -1500 (receiver range from -1500,
1500) to 1500 (receivers range from 1500, 4500). The middle figures represent the
move-out corrected CFP gather (= image gather) with a source sampling of 30 m.
The trace on the left-hand side is the one-way image trace obtained with a source
sampling of 15 m. The one-way image trace at the right-hand side is obtained with
a source sampling of 150 m.
Due to the acquisition geometry the shot records with a position x > 0 do not have a
stationary phase ray which is measured by the receivers. The missing Fresnel zone is
observed in the vanishing amplitude of the focus point response shown in figure 5.31d
for positions where x > 0. In 5.31a the receiver spacing is 600 m and the individual
contributions of the different offsets are clearly visible. In (d) the different offsets are
destructively interfering at t > 0.4 s which leaves only the undisturbed focus point
response at t = 0.4. The image trace is constructed from the move-out corrected
CFP gathers by an addition of all the traces. Using a dense source sampling (15
m) gives for all the different receiver spacings a correct image trace shown in the
Chapter 5: CFP technology
95
traces on the left-hand side. A sparse source sampling (150 m) gives a distorted
one-way time [s]
0
one-way time [s]
one-way time [s]
0
1.0
a. ∆xr = 600 [m]
-750
0
750
1500
0
0.5
1.0
b. ∆xr = 300 [m]
-1500
0
0
-750
0
750
1500
0
0.5
0.5
-750
0.5
0.5
Figure 5.31
0
∆xs = 150
0
0
750
1500
0
∆xs = 150
0
0.5
1.0
∆xs = 15
1.0
c. ∆xr = 150 [m]
-1500
0
0
0
∆xs = 150
0.5
1.0
∆xs = 15
0
0.5
1.0
∆xs = 15
0
1.0
1500
0.5
-1500
0
0
0.5
1.0
lateral position [m]
0
750
1.0
∆xs = 15
0
1.0
-750
0.5
0.5
1.0
one-way time [s]
-1500
0
0
d. ∆xr = 30 [m]
1.0
∆xs = 150
Common offset contributions for different source (∆xs ) and receiver (∆xr ) sampling
rates. The traces on the left and right-hand side are image traces. Note that for the
image trace it is sufficient that either the source or the receiver sampling is dense.
96
5.5 3-Dimensional CFP gathers
one-way time [s]
-1500
0
-750
0
750
1500
0.5
-1500
0
-750
0
750
1500
lateral position [m]
0
750
1500
0.5
1.0
1.0
a. ∆xs = 30 [m]
-1500
0
-750
lateral position [m]
0
750
0.5
b. ∆xs = 60 [m]
1500
-1500
0
-750
0.5
1.0
1.0
c. ∆xs = 150 [m]
Figure 5.32
d. ∆xs = 300 [m]
Combining focusing in detection and focusing in emission for the first focusing step.
The receiver sampling is fixed at 30 m and the source sampling is varying. Note that
the focusing in detection results (positioned on the source positions) are undistorted.
The result for focusing in emission (positioned on the receiver positions) gets distorted
due to the sparse source sampling.
image traces if the receivers are also sparsely sampled. However, a dense receiver
sampling and a coarse source sampling gives again a undistorted image trace. These
observations are of importance if one considers the 3D marine acquisition geometry
where the sampling in the cross-line direction is often coarse but the source sampling
is dense.
In the examples shown above the first focusing step was carried out for focusing
in detection. It is also possible to carry out the first focusing step for focusing in
emission and combine these results with the results for focusing in detection. The
results of these combined first focusing processes for different source sampling rates
were already shown in figure 5.11. In figure 5.32 the same result are shown but now
with a different source sampling rate. In this figure it is interesting to observe that
due to the dense receiver spacing, which is kept constant at 30 m for all figures,
the traces for focusing in detection (left-hand side) are always undistorted. Due to
the sparse sampling for the source positions the traces for focusing in emission are
distorted. This distortion is observed in the right-hand side of the pictures in figure
5.11.
Chapter 5: CFP technology
97
∆xs = 30 [m]
∆xr = 30 [m]
c = 2000 [m/s]
x
y
∆yr
∆xr
cross-line
z
⊗
Figure 5.33
5.5.2
in
e
-lin
3-dimensional acquisition geometry which is used to explain some aspects of the construction of 3D CFP gathers. The streamer arrays are fixed and the source is only
moving along the x = 0 streamer line.
A simple 3D data example
Now that the results of sparse sampling in the 2-dimensional case have been investigated in the previous subsection, the principles of construction a CFP gather for
3-dimensional data can be discussed. In this section only the principles and some
guidelines and pitfalls are given for the construction of a CFP gather in 3 dimensions.
The construction of 3D CFP gathers for more complicated examples is subject of
another research project carried out by John Bolte.
In figure 5.33 the subsurface model is shown which is used to explain the construction
of a 3D CFP gather. The model consists of one flat reflector at a depth of 800 m
and a velocity of 2000 m/s above the layer. The focus point is chosen at x = 0, y =
0, z = 800. The marine like acquisition geometry is also shown in figure 5.33; the
sources are moving in the x−direction from x = −1500, y = 0 to x = 1500, y = 0
with ∆xs = 30 m, the streamer lines are positioned in between y = 0 and y = −1500
m. The in-line receiver sampling ∆xr is 30 m and the cross-line receiver distance
∆yr is variable for the different experiments.
98
5.5 3-Dimensional CFP gathers
one-way time [s]
-1500
0
-750
lateral position [m]
0
750
1500
0.5
-1500
0
lateral position [m]
0
750
1500
0.5
1.0
1.0
a. streamer at yr = 0 [m]
-1500
0
one-way time [s]
-750
-750
0
b. streamer at yr = −510 [m]
750
0.5
1500
-1500
0
-750
0
750
1500
0.5
1.0
1.0
c. streamer at yr = −1020 [m]
Figure 5.34
d. streamer at yr = −1500 [m]
Focusing in detection of four different streamer arrays. The dotted line in the plot
indicates t = 0.4 [s]. Note that the streamer at yr = 0 contains the stationary phase
contribution at xs = 0.
The first experiment carried out is a focusing in detection for every streamer array
separately. For a complete first focusing result the focused streamers should be
added together to obtain one trace out of the 3D CFP gather. However, by looking
at the individual contributions of the streamers the construction of the 3D CFP
trace can be studied in more detail. In figure 5.34 four focused streamer arrays
are shown, along the horizontal axis the source position is shown. Note that the
correct time of the focus point response should occur at t = 0.4 s. This time is only
observed for the focused streamer at yr = 0 with the source positioned at xs = 0.
In the other focused streamer arrays the focus point response is not observed at this
time, meaning that the stationary phase ray is not measured by these streamers. In
figure 5.35 the synthesized streamer traces are shown for a section in the cross-line
direction. In these cross-sections the addition of the different streamer contributions
can be observed. For the densely sampled sections shown in figure 5.35 the addition
of the different streamers will give an undistorted CFP trace. However, by taking a
sparse sampling in the y−direction the individual contributions of the streamers will
become visible in the CFP trace. This effect is similar to the 2D CFP trace which
was built with sparsely sampled receivers as discussed in the previous section.
The effect of sparsely sampled streamer arrays in the cross-line direction is shown
in figure 5.36. The pictures presented in figure 5.36 are in fact 3D CFP gathers for
one source line; the synthesis is carried out for all receivers belonging to one shot
Chapter 5: CFP technology
one-way time [s]
-1500
0
lateral position [m]
-750
0
0.5
-1500
0
lateral position [m]
-750
0
0.5
1.0
1.0
a. cross-line at xs = 0 [m]
-1500
0
one-way time [s]
99
-750
0.5
b. cross-line at xs = −510 [m]
0
-1500
0
-750
0
0.5
1.0
1.0
c. cross-line at xs = −1020 [m]
Figure 5.35
d. cross-line at xs = −1500 [m]
Focusing in detection of four different streamer arrays. The dotted line in the plot
indicates t = 0.4 [s]. Note that the cross-section for xs = 0 contains the stationary
phase trace. All traces are plotted on the same scale.
position. The situation shown in figure 5.36c is representative for the acquisition
geometry for marine data. Note that due to the sparse sampling in the y−direction
the individual contributions of the streamers become visible in the CFP gather.
Building an CFP image trace with only the source contributions from one y−level
(as shown in figure 5.36) will give a poor image. However, according to the discussion
from the previous section the quality of the image trace can be saved if the source
sampling in the y−direction is dense. The image trace for two different sampling
rates in the y−direction and several streamer distributions in the y−direction are
shown, together with the move-out corrected CFP gathers in figure 5.37. Note the
similarity with figure 5.31.
From figure 5.37 and the foregoing examples it is concluded that:
• For focusing in detection the shot sampling can be coarse but the receiver
sampling must be fine.
• For focusing in emission the receiver sampling can be coarse but the shot
sampling must be fine.
• The Fresnel stack in 3D CFP gathers is distorted in the coarse sampled direction.
• To combine the contributions into an alias free image either the shot or receiver
sampling must be fine for both the in-line and cross-line direction.
100
5.5 3-Dimensional CFP gathers
-750
lateral position [m]
0
750
1500
-1500
0
one-way time [s]
one-way time [s]
-1500
0
0.5
1.0
lateral position [m]
0
750
1500
0.5
1.0
a. ∆yr = 30 [m]
-1500
0
one-way time [s]
-750
-750
0.5
b. ∆yr = 90 [m]
0
750
1500
-1500
0
-750
0
750
1500
0.5
1.0
1.0
c. ∆yr = 150 [m]
Figure 5.36
d. ∆yr = 300 [m]
Addition of the individual streamers for different distances between the streamer lines.
Note that for a streamer spacing of 150 [m] the contribution of the individual streamers
can still be recognized.
For the construction of a correct image gather it is important that the Fresnel zone
is sampled properly. If the sampling in the x or y direction is coarse it means that a
regularization algorithm must be used to interpolate the data in the coarsely sampled
direction in such a way that the image trace can be built up without the aliasing
effects. How the CFP technology can be used to regularize the data is explained in
the next subsection.
In the foregoing example the response of a flat reflector was considered, which means
that for a focus point positioned within the acquisition geometry the Fresnel zone
will always be measured by one of the streamers. For a dipping reflector it is argued,
from the examples shown for the 2-dimensional situation, that the Fresnel zone is
easily shifted outside the acquisition aperture. For the 3-dimensional situation this
means that if a layer is dipping in the cross-line direction the streamers arrays (with
a typical maximum width of 500 m) in the y−direction will easily ’miss’ the Fresnel
zone for a focus point positioned within the acquisition geometry. In that case the
different contributions of the streamers will not give a constructive contribution to
the image trace and interpolation will not help either. The desired acquisition width
as function of the dip and depth of the reflector is given by width = z tan (2α) + af ,
where z is the depth of the focus on the reflector positioned below the source position,
α the dip of the layer and af the width of the Fresnel zone. Note that a solution
to this problem can be found by shifting the focus point until one of the streamers
Chapter 5: CFP technology
0
one-way time [s]
0
0
one-way time [s]
one-way time [s]
-1500
0
-750
0
0
0
1.0
b. ∆yr = 300 [m]
-1500
0
-750
0
0
0.5
0.5
0
0
-1500
0
Figure 5.37
-750
0
0
∆ys = 150
0
∆ys = 150
0
0.5
1.0
∆ys = 15
1.0
c. ∆yr = 150 [m]
0.5
0.5
0
0.5
1.0
∆ys = 15
∆ys = 150
0.5
1.0
0
1.0
1.0
a. ∆yr = 600 [m]
∆ys = 15
0
0.5
0.5
0.5
1.0
0
1.0
∆ys = 15
0
1.0
0
0.5
0.5
1.0
one-way time [s]
0
lateral position [m]
-1500
-750
101
d. ∆yr = 30 [m]
1.0
∆ys = 150
Common offset contributions for different source (∆ys ) and receiver (∆yr ) sampling
rates where xs (= 0) is kept constant. The traces on the left and right-hand side are
image traces. Note that for the image trace it is sufficient that either the source or the
receiver sampling is dense.
102
5.5 3-Dimensional CFP gathers
detects the Fresnel zone of the defined focus point. The Fresnel zone is characterized
by a (local) minimum arrival time of the focus point response.
The influence of feathering, irregular shooting patterns and land acquisition geometries are not treated in this thesis. For a detailed discussion about irregular sampling
and different acquisition geometries in 3D data the reader is referred to the work of
Koek et al. (1996); Koek (1997). Choosing an optimum acquisition geometry given
a certain focus point illumination can be investigated by using the focusing beams
as explained in section 5.4. The resolution at the focus point is defined by the source
and receiver patterns at the surface. The focusing beam for the imaging result is
defined by the multiplication of the beam for focusing in detection with the beam
for focusing in detection. The resolution beam for two streamers is determined by
addition of the beams for focusing in detection for both streamer arrays (for more
about resolution aspects see Berkhout, 1984; Wapenaar, 1997).
5.5.3
Regularization of coarsely sampled data
Due to the coarse sampling in the cross-line direction the Fresnel zone is inadequately sampled which gives rise to a bad image quality. Therefore a regularization
algorithm based on the CFP technology can be helpful to avoid a bad construction of the image trace. The CFP gather is an excellent domain for interpolation,
because all events in the CFP gather belong to the same point in the subsurface.
The input of the CFP based regularization algorithm is a coarsely sampled CFP
gather and a finely sampled focusing operator. The first step in the regularization
scheme is a 2-dimensional cross-correlation between the (coarsely sampled) CFP
gather with the (finely sampled) focusing operator. The resulting bifocal image is
tapered around zero-time and zero-spatial lag to remove the aliased energy which
is positioned around its center. The next step, a 2-dimensional convolution of the
filtered bifocal image with the finely sampled focusing operator, will restore the CFP
gather on the finely sampled grid of the focusing operator. The obtained interpolated
traces are inserted in the original coarsely sampled CFP gather to avoid distortion
of the original traces. Note that a second iteration can improve the result of the
interpolated traces even more.
An illustration of the proposed regularization scheme is given in figure 5.38. The
input CFP gather represents the focus point response for focusing in detection of a
point at 800 m depth on a dipping reflector (20◦ ). The resulting CFP gather (a)
is sampled with the source sampling rate of 120 m. The focusing operator (b) is
sampled at 30 m. The 2-dimensional cross-correlation is shown in figure 5.38c, the
aliased energy is positioned around the bifocal result. Removing the aliased energy
from the bifocal result and convolution with the fine sampled focusing operator gives
a reconstruction of the CFP gather on the sampling rate of the focusing operator.
The traces missing in the coarse sampled CFP gather are now filled in by the traces
Chapter 5: CFP technology
103
obtained with the result after convolution (f). The result after a second iteration of
one-way time [s]
-1500
0
-750
lateral position [m]
0
750
1500
0.5
-1500
0
-1500
-0.5
one-way time [s]
1500
1.0
a. coarse sampled CFP gather
-750
0
750
b. fine sampled focusing operator
1500
-1500
-0.5
0
0
0.5
0.5
1.0
1.0
1.5
c. bifocal image with aliased energy
-1500
0
one-way time [s]
lateral position [m]
0
750
0.5
1.0
1.5
-750
-750
0
750
1500
0.5
-750
0
750
1500
750
1500
d. tapered bifocal image
-1500
0
-750
0
0.5
1.0
1.0
e. 2 dimensional convolutuon with (b)
-1500
0
-750
0
750
1500
0.5
f. regularization result one iteration
-1500
0
-750
0
750
1500
0.5
1.0
1.0
g. regularization result two iterations
Figure 5.38
h. fine sampled CFP gather
Regularization of a CFP gather with the scheme as discussed in the text. Due to the
tapering (d) of the bifocal result (c) the aliased energy is removed and convolving with
the focusing operator gives the interpolated traces (e). In (h) a fine sampled CFP gather
is shown for comparison.
104
5.5 3-Dimensional CFP gathers
depth [m]
-1500
0
lateral position [m]
0
750
1500
1000
2000
2000
a. sampling ∆x = 15 [m]
-750
0
1000
2000
-1500
0
-750
lateral position [m]
0
750
1500
1000
-1500
0
depth [m]
-750
750
1500
b. sampling ∆x = 60 [m]
-1500
0
-750
0
750
1500
1000
c. sampling ∆x = 150 [m]
Figure 5.39
2000
d. irregular sampling ∆x = 60 ± 30 [m]
Focusing wavefronts for different acquisition patterns at the surface. Note the difference
between the wavefronts above and below the focus point at z = 1000 m.
the same procedure is shown in figure 5.38g. The CFP gather calculated with a fine
source sampling rate is shown in (h). Comparing figure 5.38h with the result shown
in figure 5.38g it is observed that for the offset close to the edge the interpolated
traces have a slightly lower amplitude than the traces from the CFP gather, more
iteration will probably solve this problem. The traces at the small CFP offsets fit
perfectly within the original traces. Note that the finite aperture artifact in the right
upper side of the CFP gathers is not reconstructed very well.
The regularization algorithm is based on the principle shown in figure 5.39. Figure
5.39 shows how wavefronts, traveling in a homogeneous medium, are focused for
different acquisition patterns at the surface. For a coarse sampling interval of ∆x =
60 m as shown in (b), or ∆x = 150 m as shown in (c), between the traces at the
surface it is observed that the wavefront below its focus point is less disturbed by
the coarse sampling. During the propagation the wavefront gets reconstructed. For
the irregular data set, shown in (d), the same kind of reconstruction is observed.
This reconstruction property can also be used in the estimation of weathered layer
properties and will be discussed briefly in chapter 7. Note that the focus at depth
Chapter 5: CFP technology
105
level z = 1000 m has the same sharp resolution for all the used examples. The
output of the regularization algorithm is in fact the measurement of the wavefront
at z = 2000 m after an lateral selection around the focused wavefront at z = 1000
m. The topic of regularization in the CFP domain is an important subject in the
data acquisition programme of the DOLPHIN consortium (Berkhout and Thorbecke,
1996).
5.6
New developments
The applications of CFP gathers are still growing, new techniques which are based
on CFP gathers are developed and tested. In chapter 7 and 8 some of these new
techniques are shown on field and synthetic data, in this section the applications of
CFP gathers which might become important in the near future are briefly discussed
to make the overview of the possibilities of CFP gathers complete.
Target focusing; focusing of CFP gathers
By choosing several focus points close to each other, for example on the same lateral
position but at different depth levels, and making use of the interpretation that a
CFP gather for focusing in emission is the response of a source at the focus point and
receivers at the surface, a second focusing step can be applied on the CFP gathers.
In this way it is possible to collect the data from a specific point in the subsurface
which is difficult to detect in the surface data. In figure 5.40 an application of this
idea is sketched; close to a flank of a salt dome focus points are chosen at several
depth levels (the gray dots). After the calculation of the CFP gathers for focusing
in emission a second focusing step is used to focus the focus point responses on a
point positioned on the flank of the salt dome (the black dot). Application of this
idea can be useful if the macro model is already accurate and one want to image the
flank of the reflector better. Note that in this method only the direct reflections of
the flank are used and not the second order reflections.
Krebs et al. (1996) have shown that the use of accurate migration operator can
improve the quality of the image of the flank of a salt dome. In the paper accurate
migration operators are obtained by measuring the focus point responses, from a
well close to the flank of the salt dome, at the surface. In the CFP method accurate
operators can be obtained by using the operator updating mechanism which will be
discussed in the next chapter.
CFP imaging of elastic data
CFP imaging of elastic data can be done in two ways. One can first apply a decomposition into P − and S−wave potentials and then carry out CFP imaging per
data type using the acoustic algorithm, or one can integrate the decomposition and
CFP construction into one process (Berkhout et al., 1996). The integration of the
synthesis operator with the decomposition process was part of a MSc. project car-
106
Figure 5.40
5.6 New developments
Focusing of CFP gathers to illuminate the flank of a salt dome. If direct reflection
information of the flank of the salt dome is measured at the surface then it is possible to
unveil this information by a secondary focusing step.
ried out by Hondius (1993) and Hulshoff (1993) and its currently under investigation
again. Note that by using the automatic operator updating process on elastic data
one does not know if the event the operator is converging to belongs to a P − or
S−wave. If it is possible to determine to what kind of converted wave the operator
belongs to one can build P − and S−wave one-way image gathers and by using the
operator inversion scheme it is also possible to estimate a P − and S−wave macro
model.
Integration of CFP technology and multiple removal
The successful multiple removal scheme of Verschuur (1991) was originally developed
for shot records. However, it has recently been shown (Berkhout and Verschuur,
1996b) that the same multiple removal scheme also works on CFP gathers without
the need for any changes to the multiple removal algorithm. An advantage of the
CFP approach is that it is easier to apply the multiple removal scheme on 3D data.
Another advantage is that the multiple removal of interbed multiples can be combined with the imaging. By making use of bifocal images it is also possible to image
the energy of the multiples as shown in Berkhout and Verschuur (1996a).
Chapter 6
Operator updating
The influence of erroneous focusing operators on the constructed CFP gather is of
great importance, because by analyzing the synthesis process an important link with
the correct focusing operator can be found. To describe how the CFP gather is built
with an erroneous focusing operator some additional theory is needed. To describe
this theory the matrix notation introduced in chapter 4 is used.
The focusing operator for focusing in detection is defined as (after equation (4.31))
∗ −1
Fi− (zm , zr ) = Ii− (zm ) W+ (zm , zr ) D− (zr )
(6.1)
with Ii− (zm ) a unit row vector with a 1 at the ith position at depth zm and Fi− (zm , zr )
the focusing operator acting at the receiver positions at the surface.
Using equation (6.1) into the WRW representation of equation (4.30) gives an expression of the data after focusing of the detector array
Pi− (zm , zs ) = Ii− (zm )R+ (zm )W+ (zm , zs )D+ (zs )S(zs ),
(6.2)
where equation (6.2) is an expression for the CFP gather for focusing in detection.
∗
Note that if [W+ (zm , zr )] represents correct propagation in the background model
then the result of the first focusing step Pi− (zm , zs ), represented by equation (6.2),
is in traveltime equal to the time reversed focusing operator Fj+ (zs , zm ), represented by equation (4.32). To investigate the influence of erroneous operators on
the CFP gather our only interest is in traveltimes. Therefore the following assumptions are made; the source and receiver directivity operators D± are represented by unit matrices, the source function is assumed to be independent of the
source position, S(zs ) = IS0 . Note that if the sources and receiver occupy the
same positions in space (indicated with z0 ) the focusing operators are related by
T
Fi+ (z0 , zm ) = Fi− (zm , z0 ) . With these assumptions equation (6.1) and (6.2) re-
108
duce to
∗
Fi− (zm , z0 ) = Ii− (zm ) W̄+ (zm , z0 ) ,
Pi− (zm , z0 )
=
Ii− (zm )R+ (zm )W+ (zm , z0 )S0 ,
(6.3)
(6.4)
where a distinction is made between propagation in the macro model W̄+ (zm , z0 ) and
propagation in the true model W+ (zm , zs ). Equation (6.3), the focusing operator,
and equation (6.4), the CFP gather, show the ’principle of equal traveltime’ in its
most simple form, if the propagation matrices are equal, W̄+ (zm , zr ) = W+ (zm , zs ),
then the time-reversed focusing operator is in traveltime equal with the CFP response. Suppose that the macro model is not correct and that the true propagation
matrix can be represented by
W− (z0 , zm ) = W̄− (z0 , zm )∆W(zm ),
(6.5)
where ∆W(zm ) is the propagation effect of the macro model error. Then the time
reversed focusing operator is defined as
−
∗
∗
Fi (zm , z0 ) = Ii− (zm ) [∆W(zm )] W+ (zm , z0 ),
(6.6)
+
∗
∗
where W̄ (zm , z0 ) in equation is replaced by ∆W(zm ) [W+ (zm , z0 )] , and the CFP
response as
Pi− (zm , z0 ) = Ii− (zm )∆W(zm )R+ (zm )W+ (zm , z0 )S0 .
(6.7)
Assuming that ∆W(zm ) and R+ (zm ) are Toeplitz matrices, which means that the reflection and propagation properties don’t change with respect to the lateral position
in the area of interest, they can be interchanged, which results in
Pi− (zm , z0 ) = Ri+ (zm )∆W(zm )W+ (zm , z0 )S0 ,
(6.8)
where Ri+ (zm ) is a row vector representing the it h row of the reflection matrix
R+ (zm ).
From equations (6.6) and (6.8) it follows that the propagation errors ∆W(zm ) in
the time reversed focusing operator and the related focus point response are equal
in magnitude and opposite in phase. This is not a surprise because the propagation
time which is subtracted too much in the first focusing step should be too little in the
CFP response since the total propagation time of the data is fixed (Berkhout, 1996b).
After the construction of the CFP gather the arrival times of the erroneous focusing
operator can be compared with the response in the CFP gather. As indicated by
equation (6.6) and (6.8) the principle of equal traveltime is no longer valid for the
erroneous operator and the related focus point response. The macro model and/or
the operator must be updated to obtain a correct operator which obeys the principle
of equal traveltime.
Chapter 6: Operator updating
one-way time [s]
-1500
0
one-way time [s]
1500
0.5
1.0
1.0
a. c = 2000 z = 800
-750
0
750
1500
0.5
1.0
lateral position [m]
-750
0
750
1500
1.0
b. c = 2000 z = 640
-1500
0
1.0
d. c = 1800 z = 800
-750
0
750
1500
0.5
-750
0
750
1500
Figure 6.1
1500
c. c = 2000 z = 960
-1500
0
1.0
e. c = 1800 z = 640
-1500
0
1.0
lateral position [m]
-750
0
750
-750
0
750
1500
0.5
-750
0
750
0.5
g. c = 2200 z = 800
-1500
0
0.5
0.5
-1500
0
1.0
-1500
0
0.5
-1500
0
one-way time [s]
lateral position [m]
-750
0
750
109
1500
f. c = 1800 z = 960
-1500
0
-750
0
750
1500
0.5
h. c = 2200 z = 640
1.0
i. c = 2200 z = 960
CFP corrected shot record with different erroneous focusing operators. The shot record
used is the 50th shot (x = −750) in a fixed acquisition geometry and one flat reflector
at 800 m depth. The upper left picture (a) gives the correct CFP corrected shot record.
Note that the integration along the receiver position gives one trace in the CFP gather.
In the following sections the influence of an erroneous focusing operator in the first
and the second focusing step will be discussed. An attempt is made to detect from
the erroneous CFP gather an update formula for the macro model where the initial
focusing operator is modeled in. Finally a proposal is made where the erroneous
focusing operators are updated to a correct focusing operator.
6.1
First focusing step
Using an erroneous focusing operator will give a different construction of the CFP
gather. This is shown in figure 6.1 for several errors in the focusing operator applied
on one shot record with a source position at x = −750 (the 50th out of a range of 201
shot records) from a 1-dimensional medium with a fixed acquisition spread ranging
from −1500 to 1500 m. The reflector in the 1-dimensional medium is positioned
at 800 m depth and the velocity of the first layer is 2000 m/s. The CFP gathers
constructed with the erroneous synthesis operators are shown in figure 6.3. Note
that the trace at the lateral position x = −750 of the CFP gather in figure 6.3 is
the summation along the traces of figure 6.1. The introduced errors in the focusing
operators give CFP gathers which are not time coincident with the times defined
by the operator (indicated by the spiky event in the CFP gathers of figure 6.3). In
110
one-way time [s]
-1500
0
one-way time [s]
lateral position [m]
-750
0
750
1500
0.5
1.0
1.0
a. c = 2000 z = 800
-750
0
750
1500
0.5
1.0
lateral position [m]
-750
0
750
1500
1.0
b. c = 2000 z = 640
-1500
0
1.0
d. c = 1800 z = 800
-750
0
750
0.5
1500
-750
0
750
1500
Figure 6.2
1500
c. c = 2000 z = 960
-1500
0
1.0
e. c = 1800 z = 640
-1500
0
1.0
lateral position [m]
-750
0
750
-750
0
750
1500
0.5
-750
0
750
0.5
g. c = 2200 z = 800
-1500
0
0.5
0.5
-1500
0
1.0
-1500
0
0.5
-1500
0
one-way time [s]
6.1 First focusing step
1500
f. c = 1800 z = 960
-1500
0
-750
0
750
1500
0.5
h. c = 2200 z = 640
1.0
i. c = 2200 z = 960
CFP corrected shot record with different erroneous focusing operators. The shot record
used is the 1rst shot (x = −1500) in a fixed acquisition geometry ranging from −1500
to 1500 m with one flat reflector at 800 m depth.
figure 6.1, where the correct time of the stationary phase contribution is indicated
by the dotted line, it is observed that the Fresnel zone is shifted in time and space,
with respect to the correct position (a), due to the error in the operator. This shift
in the Fresnel zone disturbs the integration result rigorously when the Fresnel zone
is shifted outside the aperture range. In figure 6.2 the same erroneous operators are
used but now for a shot with a source position at x = −1500, which gives the trace at
x = −1500 in the CFP gathers shown in figure 6.3. This shot record is positioned at
the edge of the model and the Fresnel zone can be easily shifted outside the aperture
range due to an erroneous operator. For example in figure 6.2i the Fresnel zone is
shifted outside the aperture and the contribution in the CFP gather at x = −1500
of figure 6.3i cannot be used for a macro model or operator update. Note that the
focusing operators with a positive velocity error can be interpreted as a high angle
filter which cuts out the higher angles in the CFP gather response as shown in figures
6.3g, h and i.
If the Fresnel zone for an erroneous operator stays within the aperture range then
the CFP gather can be corrected for the erroneous operator and the calculation of a
new CFP gather with an updated operator is not necessary. It can be advantageous
to look at methods which keep the Fresnel zone within the aperture range. In figure
6.2g, h and i it is observed that the Fresnel zone is shifted outside the aperture
range due to the error in the operator. The first synthesis process may be improved
Chapter 6: Operator updating
one-way time [s]
-1500
0
one-way time [s]
1500
0.5
1.0
1.0
a. c = 2000 z = 800
-750
0
750
1500
0.5
1.0
lateral position [m]
-750
0
750
1500
1.0
b. c = 2000 z = 640
-1500
0
1.0
d. c = 1800 z = 800
-750
0
750
0.5
1500
-750
0
750
1500
Figure 6.3
1500
c. c = 2000 z = 960
-1500
0
1.0
e. c = 1800 z = 640
-1500
0
1.0
lateral position [m]
-750
0
750
-750
0
750
1500
0.5
-750
0
750
0.5
g. c = 2200 z = 800
-1500
0
0.5
0.5
-1500
0
1.0
-1500
0
0.5
-1500
0
one-way time [s]
lateral position [m]
-750
0
750
111
1500
f. c = 1800 z = 960
-1500
0
-750
0
750
1500
0.5
h. c = 2200 z = 640
1.0
i. c = 2200 z = 960
CFP gathers for a flat reflector with different erroneous operators. The shot records
are simulated in a fixed acquisition geometry. The spiky event in the gather indicate the
traveltimes of the used focusing operator.
if not only one trace for the CFP gather is calculated, but also neighboring traces
of different CFP gathers, constructed with a lateral shifted operator, are used. In
figure 6.4 the shot record with a source at x = −1500, corrected with the erroneous
synthesis operator which was also used in 6.2i, is shown for different lateral shifts in
the synthesis operators. In these results it is observed that the Fresnel zone, which
was initially shifted outside the aperture range due to an erroneous operator, can
be included in the CFP gather by shifting the focusing operator laterally and using
these results in the construction of the CFP gather. Note that the laterally shifted
synthesis operators, which defines another focus point, are able to bring the Fresnel
zone back in the aperture range at a different time with a different amplitude after
integration. Note also that in considering all possible laterally shifted operators
within the aperture range means that a 2-dimensional cross-correlation is carried
out between the shot record and the synthesis operator. This 2-dimensional crosscorrelation can also be used to built up a CFP gather more efficiently.
In figure 6.5a, b and c the 2-dimensional cross-correlation between the correct operator used in figure 6.2a and three shot records at positions x = −1500, x = −750
and x = 0 is shown. The trace which is also present in the CFP gather of figure
6.3a is positioned at respectively x = −1500, x = −750 and x = 0. Note that
for a 1-dimensional medium a 2-dimensional cross-correlation between the operator
and a shot record can be interpreted as building a CFP gather with different shot
112
-1500
0
6.1 First focusing step
-750
0
750
1500
0.5
1.0
-750
0
750
1500
0.5
1.0
a. shift of x = 250
-1500
0
-750
0
750
1500
0.5
1.0
-1500
0
1.0
b. shift of x = 0
-1500
0
1.0
Figure 6.4
-750
0
750
1500
0.5
-750
0
750
1500
0.5
d. shift of x = −500
-1500
0
c. shift of x = −250
-1500
0
-750
0
750
1500
0.5
1.0
e. shift of x = −750
f. shift of x = −1000
1-dimensional cross-correlation between laterally shifted erroneous focusing operators
(c = 2200 z = 960) and one shot record with a source at x = −1500.
records where the spatial sampling rate of the shot positions is given by the spatial
sampling rate of the focusing operator. The 2-dimensional cross-correlation gives
for a 1-dimensional medium exactly the same result (beside the finite aperture artifacts) as the full synthesis process in the spatial domain using all shot records. In
combining these 2-dimensional cross-correlation results in the construction of the
CFP gather, the lateral position of the traces in the cross-correlation result must
be mirrored around the source position. Or in other words; the synthesis result of
a shifted operator and a shot record is in a 1-dimensional medium equivalent with
the synthesis of the non-shifted operator and a shifted shot record, where the shift
of the shot record is in the other direction than the shift of the synthesis operator.
In figure 6.5d, e and f three CFP gathers are shown which are constructed by using
only a few shot records and the 2-dimensional cross-correlation between the correct
-1500
0
-750
0
750
1500
0.5
1.0
-750
0
750
1500
0.5
1.0
a. shot at x = −1500
-1500
0
-750
0
750
1500
0.5
1.0
-1500
0
Figure 6.5
1.0
b. shot at x = −750
-1500
0
1.0
-750
0
750
1500
750
1500
0.5
-750
0
750
0.5
d. shot at every 1500 m
-1500
0
1500
c. shot at x = 0
-1500
0
-750
0
0.5
e. shot at every 750 m
1.0
f. shot at every 500 m
2-Dimensional cross-correlation between the correct focusing operator and three shot
records (a, b and c). Addition of a limited number of cross-correlated shot records is
shown in d, e and f.
Chapter 6: Operator updating
-1500
0
-750
0
750
1500
0.5
1.0
-750
0
750
1500
0.5
1.0
a. shot at x = 0
-1500
0
-750
0
750
0.5
1.0
-1500
0
1500
Figure 6.6
1.0
b. shot at x = −255
-1500
0
1.0
-1500
0
-750
0
750
1500
0.5
-750
0
750
0.5
d. shot at every 300 m
113
1500
c. shot at x = −495
-1500
0
-750
0
750
1500
0.5
e. shot at every 150 m
1.0
f. shot at every 75 m
2-Dimensional cross-correlation between the correct operator and three shot records (a,
b and c) from the syncline model. Making use of a limited number of shot records and
the 2-dimensional cross-correlation an image can be built up as shown in d,e and f.
operator and these selected shot records.
Using the 2-dimensional cross-correlation to build up the CFP gather is only possible
for the 1-dimensional medium as presented in figure 6.5, because in a 1-dimensional
medium a laterally shifted shot record represents a correct shot record. In the
syncline model with a focus point defined at the deepest point of the syncline this
is not the case. In figure 6.6 the 2-dimensional cross-correlation is shown for three
shot records with source positions at x = 0, x = −255 and x = −495 m. In
this experiment it is known that due to the lateral variation in the medium the 2dimensional cross-correlation gives only a correct contribution in the CFP gather for
the zero-lag trace, the laterally shifted operators are not correct anymore. However,
in figure 6.6 it can be observed that the traces close to the zero lag trace are still
useful, further away from the zero-lag trace the focusing breaks down and is not
useful anymore. In figure 6.6d, e and f CFP gathers are shown which are constructed
with the 2-dimensional cross-correlation by using a selection of all available shot
records. The ’missing’ shot positions are obtained from the 2-dimensional crosscorrelation.
In conclusion; the 2-dimensional cross-correlation result can be used to built up
the CFP gather if the assumption that the medium is laterally invariant is valid. If
the medium is strongly laterally variant only the zero-lag trace can be used (which
in fact is done in the full synthesis process by using all shot records). If there is a
big gap between two shots records the 2-dimensional cross-correlation between the
focusing operator and the shot record can be used as an interpolation operator for
the gap between the shots. The ’missing’ shot in the gap is created by copying the
neighboring shot and changing the source coordinate, which is obviously wrong for
a non 1-dimensional medium.
114
6.2 Searching for updating formulas for a flat reflector
xs
z
Figure 6.7
6.2
shot record
xr
c
xr
focusing operator
z + ∆z
xr
c + ∆c
Ray paths for a shot record of a flat reflector and an (erroneous) focusing operator.
Searching for updating formulas for a flat reflector
In this section an attempt is made to derive analytical expressions for the curves
observed in the CFP corrected shot record for a flat reflector in the subsurface. In
the analysis only the arrival times of the reflections are taken into account. The well
known traveltime curve for a flat reflector at depth z with a source at position xs
and a receiver at xr is described by
p
4z 2 + (xr − xs )2
Td (xs , xr ) =
,
(6.9)
c
where the subscript d in Td refers to the data measured at the receiver positions xr
at the surface and c is the velocity of the medium above the reflector. The traveltime
curve for a time-reversed focusing operator with a focusing point in the subsurface
is given by
p
(z + ∆z)2 + (xr − x)2
,
(6.10)
T̄s (x, xr ) =
c + ∆c
where the subscript s in T̄s refers to the synthesis operator (see also figure 6.7) and
∆z and ∆c represent the depth and velocity error with respect to the true model.
Application of the focusing operator to a common shot gather (fixed xs ) can be
interpreted as a time convolution between the traces of the data and the traces of
the focusing operator. In terms of traveltimes this time convolution is represented
by
p
p
4z 2 + (xr − xs )2
(z + ∆z)2 + (xr − x)2
Td (xs , xr ) − T̄s (x, xr ) =
−
. (6.11)
c
c + ∆c
From the two-way traveltime from source to reflector and back to receiver, the oneway traveltime, from reflector to receiver, is peeled off by the synthesis operator.
The synthesis process itself is an integration over all the convolved traces of the
CFP corrected common shot gather (see also equation (5.1) so,
Z
A(xr , xs , x) exp (jωΦ(xr , xs , x))dxr ,
(6.12)
P (x, xs ) =
∂D
Chapter 6: Operator updating
115
where A is an amplitude factor and Φ represents the traveltime function defined
by equation (6.11). Equation (6.12) gives the synthesized trace, in the frequency
domain, for one common shot gather. Repeating the process for all common shot
gathers will give the CFP gather for focusing in detection. The stationary phase
approximation simplifies the integration in equation (6.12) to the evaluation of the
stationary points with respect to xr . The stationary points of equation (6.11) are
given by
∂
Φ(xr , xs , x) = 0
xr
(xr − x)
(xr − xs )
p
p
−
=0
2
2
c 4z + (xr − xs )
(c + ∆c) (z + ∆z)2 + (xr − x)2
(6.13)
which solution is equivalent to
p
p
(xr − xs )(c + ∆c) (z + ∆z)2 + (xr − x)2 − (xr − x)c 4z 2 + (xr − xs )2 = 0
(6.14)
Solving equation (6.14) for xr in a closed form is not trivial, but by considering
only a depth or velocity error for a focus point at x = 0 it is possible to derive, by
inspection and numerical evaluation, some properties of the stationary point and the
CFP corrected shot record. Having found the stationary point it is possible to replace
the integration in equation (6.12) by the stationary phase solution (Bleistein, 1984).
The stationary phase solution for the traveltimes can be calculated by substituting
the stationary point in equation (6.11). By doing this calculation for all common
shot gathers (variable xs ) the traveltime curve for the CFP gather is found.
6.2.1
Zero depth and zero velocity error (∆z = 0 and ∆c = 0)
Using the correct macro model will give for the CFP gather exactly the same hyperbola as the focusing operator. For the focus point at x = 0 equation (6.14) reduces
then to
p
p
(xr − xs )c z 2 + x2r − xr c 4z 2 + (xr − xs )2 = 0
(6.15)
which has the simple solution xr = −xs . Substituting this solution (stationary
point) in equation (6.11) for all shot positions gives for the CFP gather
Tcf p (x = 0, xs ) =Td (xs , xr = −xs ) − T̄s (x = 0, xr = −xs )
p
z 2 + x2s
=
,
c
(6.16)
with the synthesis operator
p
z 2 + x2s
T̄s (x = 0, xs ) =
.
c
Note that equation (6.17) is indeed the same as equation (6.16).
(6.17)
116
6.2 Searching for updating formulas for a flat reflector
positive depth error
negative depth error
Tcf p
T̄s
T̄s
z
z
+∆z
Figure 6.8
6.2.2
Tcf p
−∆z
Depth errors in the focusing operator: depth too deep (a) and depth too shallow (b)
gives different CFP gathers, which can be interpreted as the response of a virtual source
above (a) or below (b) the reflector.
Depth errors (∆z 6= 0 and ∆c = 0)
In the case of a depth error and a zero velocity error equation (6.14) reduces to
p
p
(xr − xs )c (z + ∆z)2 + x2r − xr c 4z 2 + (xr − xs )2 = 0.
(6.18)
The solution for the stationary points in this equation can still be expressed in a
closed form and is given by
∆z + z
.
(6.19)
xr = xs
∆z − z
Substitution of this solution in equation (6.11) gives for the CFP gather
p
(z − ∆z)2 + x2s
Tcf p (x = 0, xs ) =
,
c
(6.20)
where the synthesis operator is given by
T̄s (x = 0, xs ) =
p
(z + ∆z)2 + x2s
.
c
(6.21)
From equation (6.21) it is observed that the time correction, to the downgoing source
wave field at the surface, with the synthesis operator is for a positive ∆z too large.
The resulting CFP gather represented by equation (6.20) can be interpreted as the
response of a virtual source positioned ∆z m above the reflector (see left picture
in figure 6.8). For a negative depth error the time correction with the synthesis
operator is too small and the CFP gather can be interpreted as the response of a
virtual source below the reflector position (or alternatively as the reflected response
from a source above the reflector, see right picture in figure 6.8).
With the derived solution of equation (6.19) it is possible to arrive at an analytical
updating formula for the macro model, by using equations (6.20) and (6.21) and
solving for the two unknowns z and ∆z.
Chapter 6: Operator updating
6.2.3
117
Velocity errors (∆z = 0 and ∆c 6= 0)
If there is only a velocity error in the synthesis operator equation (6.14) reduces to
p
p
(6.22)
(xr − xs )(c + ∆c) z 2 + x2r − xr c 4z 2 + (xr − xs )2 = 0
Unfortunately it is difficult to find a closed form solution for this equation. Therefore
equation (6.22) is solved numerically for a set of different parameters. Based on these
numerical results the following observations are made:
• If the velocity error ∆c is negative, thus using a velocity in the calculation
of the synthesis operator which is too low there is always one real numerical
solution of equation (6.22). This means that there is at least one stationary
point which gives a single contribution to the CFP gather.
• If the velocity error ∆c is positive, thus using a velocity in the calculation
of the synthesis operator which is too large, there are always at least two
real numerical solutions found. This means that there are two stationary
points which both give both a contribution to the CFP gather. The numerical
experiments also showed that one of the two stationary points found lays far
away from xr = 0.
6.2.4
CFP corrected shot record
The traveltime of the CFP corrected shot record prior to integration, thus before
summation over all the traces, will give more insight in the influence of erroneous
operators in the CFP gather. These traveltimes can be calculated for different
parameters by using equation (6.11). For the traveltime calculation a simple model
is chosen with only one reflector at 300 m below the surface and a velocity of 2000
m/s above the reflector. The receivers in this model are positioned from -1500 m to
1500 m with a spatial interval of 15 m. The 201 shots are positioned at all detectors
positions from the first detector position at -1500 m to the last detector position at
1500 m with a spatial interval of 15 m.
In figure 6.9 four different time sections, calculated by using equation (6.11), show
the influence of for depth and velocity errors as function of the receiver position in
the model for three different shot records. The shot gathers with their shot position
at xs = 0 show for the depth errors only one stationary point at xr = 0. The other
shot positions (xs = 450 and xs = 900) have a shifted stationary point position
which position on the xr -axis is given by equation (6.19). The time sections with
a velocity used in the focusing operator which is smaller than the true velocity
show that there exists indeed only one stationary point, although it was not (yet)
possible to derive it analytically. For a velocity higher than the true velocity it is
not possible to observe a stationary point for the shot positions at xs = 450 and
xs = 900 which makes it difficult to forecast, with the aid of the time section, what
118
6.2 Searching for updating formulas for a flat reflector
-0.6
time [s]
-0.3
0
0.3
-0.6
depth = 200 m
-0.3
xs = 0
0
xs = 450
0.3
-750
0
750
lateral position [m]
time [s]
0
-750
0
750
lateral position [m]
1500
-0.6
velocity = 1800 m/s
-0.3
xs = 0
0
xs = 450
0.3
xs = 900
0.6
1500 -1500
-0.6
-0.3
xs = 0
xs = 450
xs = 900
0.6
-1500
depth = 400 m
0.3
xs = 900
0.6
-1500
Figure 6.9
velocity = 2200 m/s
xs = 0
xs = 450
xs = 900
-750
0
750
0.6
1500 -1500
-750
0
750
1500
Time sections of three shot gathers (xs = 0, 450, 900 m) prior to integration for different
depth and velocity errors in the focusing operator (equation (6.11)). Note that for a
positive velocity error the stationary point is shifted outside the aperture.
the CFP corrected shot record will look like. It is therefore not possible to derive
a closed form expression for the updating formula’s to update a macro model. The
curves in figure 6.9 show clearly that the stationary phase point is shifted due to
erroneous operators, for large errors the point can shift outside the aperture and the
CFP trace will consists of artifacts only. This was already observed in section (6.1).
An alternative of equation (6.11), to calculate the time sections for the shot gathers
prior to synthesis, the same result can be obtained by convolving the shot gather
under consideration with the time-reversed focusing operator. The resulting record
time [s]
-1500
-0.6
-750
lateral position [m]
0
750
1500
-1500
-0.6
-0.3
-0.3
0
0
0.3
0.3
0.6
-750
lateral position [m]
0
750
1500
0.6
a. depth error = -100 m
Figure 6.10
b. velocity error = +200 m/s
Erroneous synthesis result for a shot record with xs = 450. Note the double contribution in the integrated trace with the positive velocity error (b).
Chapter 6: Operator updating
119
is shown in figure 6.10a for a shot gather at position xs = 450 and a depth error in
the synthesis operator of -100 m (depth = 200 m). In figure 6.10b the same shot
gather is shown but now with a synthesis operator with a velocity error of +200 m/s
(velocity = 2200 m/s) and the correct depth. In both pictures the same shape of
the time section as in figure 6.9 can be observed. The result of the summation over
all traces is also given in figure 6.10. The summation shows that the position of the
stationary point in figure 6.10a is indeed the only contribution in the summation.
However, in figure 6.10b two contributions are observed in the synthesized trace.
6.2.5
Move-out corrected CFP gather
The integration over all the receivers in a common shot gather must be done for all
available shot gathers to construct the CFP gather. It was already shown that if the
correct macro model is used to generate the focusing operator the principle of equal
traveltime is valid. For an erroneous macro model this principle is no longer valid.
An expression of the derivation in traveltime of the CFP gather with the erroneous
focusing operator is given by
Tmv (x, xs ) = Tcf p (x, xs ) − T̄s (x, xs ).
(6.23)
This time difference can be used to find an update for the initial macro model
which was used to model the focusing operator. If the exact traveltime curve for
Tmv is known for a certain error in the model then it is possible to compare these
analytical curves (in a least squares way) with the observed curve and determine the
velocity and depth errors made in the macro model. It has been shown that for a
depth error in the macro model it is possible to derive a closed form expression for
Tmv . In this ideal case only one update is sufficient to arrive at the correct macro
model. For a velocity, or a depth and velocity error it is not possible to derive such
a closed form. In that case numerical methods are used to calculate the stationary
points in equation (6.11) in order to determine Tmv . The scheme given in figure
6.11 calculates Tmv in a numerical way such that it can be used in the updating
scheme discussed later. The calculation of the roots of equation (6.14) is done with
next xs
c, ∆c
z, ∆z
x, xs
Figure 6.11
calculate
root(s)
calculate
Tcf p (xs )
subtract
T̄s (xs )
Tmv (xs )
Numerical calculation of Tmv for a CFP gather with a velocity and a depth error in the
focusing operator T̄s .
120
6.2 Searching for updating formulas for a flat reflector
the method of Newton-Raphson (Press et al., 1992). For negative velocity errors
and negative and positive depth errors there is only one root and the method can
be used without any problems. For positive velocity errors a-priori knowledge of the
position of one of the roots has to be built in the algorithm.
In figure 6.12 the move-out corrected CFP gathers are shown for different velocity
and depth errors in the macro model which is used to calculate the focusing operator.
If the correct model is used the move-out panel shows an event which aligns at zero
time. For a negative velocity error the move-out panels contain one simple event
(originating from the single stationary point), positive velocity errors give more
complicated move-out panels (originating from the multiple stationary points). The
positive and negative depth errors are symmetric with respect to the zero time lag in
the move-out panels and contain only one single event. The same panels can also be
found in the work of Cox (1991) who calls them Common Depth Point gathers. These
CDP gathers are obtained after shot record redatuming followed by correction with
the incident source wave field (the move-out in the focusing operator) and stacking
(which is similar to the synthesis process in the first focusing step). Cox (1991)
showed already that the CDP gather (called in this thesis move-out corrected CFP
gather) contains information which can be used to estimate the macro model.
In figure 6.13 numerically calculated curves, according to the scheme given in figure
6.11, are shown for the different velocity and depth errors used in figure 6.12. The
curves for the depth errors are calculated correctly for both positive and negative
depth errors. The curves for the velocity errors are only correct for the negative
velocity errors, the positive velocity errors show the contribution for one stationary
point only.
In figure 6.14 the move-out corrected CFP gather is shown for different combinations
of depth and velocity errors. For a positive velocity error a more complicated panel is
observed again. The results for the negative velocity errors combined with a positive
or negative depth error are used in a macro model estimation scheme which makes
use of the numerical scheme of figure 6.11.
After numerical calculation of the move-out curve Tmv this curve is compared in a
least squares manner with the tracked time section of the event in figure 6.14. The
domain, in which the search for a match with the tracked time curve is carried out,
is in depth varying from 100 to 500 m with steps of 5m and in velocity from 1600
to 2400 m/s with steps of 10 m/s. The results are shown in table 6.1 for the five
best matches. Note, that the estimated depth and velocity update gives directly the
correct answer. From table 6.1 it can be seen that the best match, based on the least
squares criterion, can be distinguished clearly from the other four nearest matches.
The best match together with the tracked time section are shown in figure 6.15.
Note the good match between the tracked time section and the calculated curve.
The foregoing examples have shown that it is possible to calculate analytical curves
Chapter 6: Operator updating
time [s]
-1500
-0.3
lateral position [m]
0
750
1500
-1500
-0.3
-0.2
-0.2
-0.1
-0.1
0
0
0.1
0.1
0.2
0.3
-0.3
time [s]
-750
0.2
c = 1600 [m/s]
0.3
-0.3
-0.2
-0.2
-0.1
-0.1
0
0
0.1
0.1
0.2
0.3
0.2
c = 1800 [m/s]
0.3
121
-750
lateral position [m]
0
750
1500
z = 100 [m]
z = 200 [m]
-0.3
time [s]
-0.2
-0.1
0
0.1
0.2
time [s]
0.3
-0.3
-0.3
-0.2
-0.2
-0.1
-0.1
0
0
0.1
0.1
0.2
0.3
-0.3
time [s]
no error
c = 2200 [m/s]
0.2
0.3
-0.3
-0.2
-0.2
-0.1
-0.1
0
0
0.1
0.1
0.2
0.3
c = 2400 [m/s]
Figure 6.12
0.2
0.3
z = 400 [m]
z = 500 [m]
Move-out corrected CFP gathers for different types of erroneous focusing operators.
Note that if the correct model is used the move-out panel shows an event which aligns
at zero time. On the left side the influence of velocity errors is shown and on the right
side the influence of depth errors is shown.
time [s]
122
6.2 Searching for updating formulas for a flat reflector
-0.3
-0.3
-0.2
-0.2
-0.1
-0.1
0
0
0.1
0.1
0.2
0.2
0.3
-1500
Figure 6.13
-750
0
750
lateral position [m]
1500
0.3
-1500
-750
0
750
lateral position [m]
1500
Move-out curves for different depth (left) and velocity (right) errors in the focusing
operator.
for the different depth and velocity errors in a single reflector model. By comparing
the tracked time section of the move-out corrected CFP gather with the calculated
curves it is possible to estimate a macro model. With this new macro model a
new focusing operator and a CFP gather can be calculated. This procedure can
be applied recursively and will terminate when the move-out corrected CFP gather
has a straight event at zero time. In that case the synthesis operator and the event
under consideration in the CFP gather are exactly the same, obeying the principle
of equal traveltime. Note that for a correct migration output horizontal alignment
is a necessary condition but not a sufficient one. In addition to horizontal alignment
the differential time with the focusing operator should be zero (Berkhout, 1996b).
time [s]
-1500
-0.3
lateral position [m]
0
750
1500
-1500
-0.3
-0.2
-0.2
-0.1
-0.1
0
0
0.1
0.1
0.2
0.3
-0.3
time [s]
-750
z = 200 [m], c = 1800 [m/s]
0.2
0.3
-0.3
-0.2
-0.2
-0.1
-0.1
0
0
0.1
0.1
0.2
0.3
z = 200 [m], c = 2200 [m/s]
Figure 6.14
0.2
0.3
-750
lateral position [m]
0
750
1500
z = 400 [m], c = 1800 [m/s]
z = 400 [m], c = 2200 [m/s]
Move-out corrected CFP gathers for combined depth and velocity errors in the focusing
operators. Note that for negative velocity errors simple curves are observed.
Chapter 6: Operator updating
initial c=1800 z=200
123
initial c=1800 z=400
velocity [m/s]
depth [m]
L2 fit
velocity [m/s]
depth [m]
L2 fit
2000
2000
2010
2010
2020
300
305
305
310
310
2.62e-04
1.52e-03
1.79e-03
2.94e-03
5.77e-03
2000
2010
2010
2020
2030
300
305
300
310
310
2.76e-04
1.33e-03
2.27e-03
4.08e-03
7.67e-03
Table 6.1 The five best matches for the time section of figure 6.14 with the negative velocity errors.
The correct model can be distinguished clearly from the other matches.
In conclusion; the move-out corrected CFP gather contains information about
the depth and velocity errors in the macro model. Using a focusing operator derived
from an initial macro model there are in general two unknowns: the velocity of the
medium, and the depth of the reflector. To find an update for an initial guess,
analytically (or numerically) calculated traveltime curves can be used to find the
best fit with the tracked curve. In the ideal case only one iteration is needed to
arrive at the correct model. But due to tracking errors and multiple arrival times in
complicated media there are in general more iterations needed.
-0.20
time [s]
-0.05
0
-0.15
0.05
z = 200 [m], c = 1800 [m/s]
0.10
-1500
Figure 6.15
6.3
-750
0
750
lateral position [m]
z = 400 [m], c = 1800 [m/s]
-0.10
1500 -1500
-750
0
750
lateral position [m]
1500
Fitted move-out curves for the combined depth and velocity errors as shown in the top
two pictures in figure 6.14.
Operator updating
In the previous section an attempt was made to derive update formula’s for depth
and velocity errors in a 1-dimensional macro model. It was shown that it is possible
to derive a closed form solution of the update formula in case of a depth error in a
one layer model. For other errors and more complicated models a numerical method
can be used to find a better macro model (Kabir, 1997). However, there is a different
approach possible which can be used to derive the correct focusing operator. Given
an erroneous focusing operator it is possible to update the focusing operator and not
124
6.3 Operator updating
the macro model. The operator is updated to any suitable event close to the initial
operator. How the correct operator can be found, from the focus point response and
the initial operator, is explained in this section for a flat and a dipping layer.
6.3.1
Flat reflector
Let us assume that the correct focusing operator for the chosen event is positioned ’in
between’ the initial operator and the event of interest (principle of equal complexity).
To illustrate the concept of ’in between’, assume a hyperbolic time behavior for the
initial synthesis operator (Ts0 ) and the focus point respons (Tcf p )
p
z̄ 2 + x2r
0
(6.24)
T̄s (xr ) =
p c̄
z 2 + x2r
Tcf p (xr ) =
(6.25)
c
with z̄ the depth of the focus point , c̄ an average (or rms) velocity and xr a receiver
position. T̄s0 (xr ) describes the time behavior of the initial synthesis operator and
Tcf p (xr ) describes the time behavior of the focus point response. The update of the
synthesis operator is given by
T̄s1 (xr ) = T̄s0 (xr ) + Tc (xr )
(6.26)
where Tc (xr ) is the update. An intuitive operator update is chosen as
Tc (xr ) =
Tcf p (xr ) − T̄s0 (xr )
,
2
(6.27)
which is the traveltime in between the operator and the focus point response at every
offset. This operator update approach gives the correct operator (using hyperbolic
operators) for xr = 0, but for larger CFP offsets the linear update is less accurate.
The accuracy at the larger offsets can be improved by making use of a second
updating step. In figure 6.16 two examples from the previous section, shown in
figure 6.3e and f, are used to demonstrate the convergence of the proposed updating
scheme. The examples contain a combined depth/velocity error in the focusing
operator for a model with one flat layer. From the results shown in figure 6.16 it is
observed that although the initial macro model and the CFP gather are different in
both examples they converge to the same answer within one iteration. The updating
of the CFP gather can be carried out in an automatic way by a time-convolution of
the traces in the CFP gather with the traces in the time-reversed focusing operator
(hence each trace is delayed with its erroneous response), followed by a renumbering
to one-way time.
In the results after the first update in figure 6.16 the correct focusing operator is
indicated by the spiky event. Careful inspection of the updated result after one iteration shows that for the higher CFP offsets the convergence to the correct operator
Chapter 6: Operator updating
one-way time [s]
-1500
0
-750
lateral position [m]
0
750
1500
0.5
-1500
0
125
-750
lateral position [m]
0
750
0.5
1.0
1.0
c = 1800 [m/s] z = 640 [m]
⇓
c = 1800 [m/s] z = 960 [m]
first iteration
-1500
0
one-way time [s]
1500
-750
0
750
0.5
⇓
first iteration
1500
-1500
0
-750
0
750
1500
0.5
1.0
1.0
result after first update
Figure 6.16
result after first update
Synthesis operator updating by taking the traveltime in between the operator and the
CFP response. Note that both CFP gathers converge very fast to the same CFP gather
with a focus point on the reflector. One iteration seems to be sufficient.
is not perfect. A second iteration will give a better convergence for the higher offsets.
However, there is also an alternative updating scheme possible which converges, for
the examples in figure 6.16, within one iteration to the correct operator.
Two-dimensional convolution (in space and time) of the erroneous time-reversed
focusing operator with its erroneous focus point response can be written in a matrixvector multiplication. For example the 2-dimensional convolution of equation (6.6),
the erroneous focusing operator, with equation (6.8) the focus point response, can
be represented by
∗
Qi− (zm , z0 ) = Pi− (zm , z0 ) F−
(6.28)
i (zm , z0 )
= Ri+ (zm )∆W(zm )W+ (zm , z0 ) [∆W(zm )]∗ W+ (zm , z0 )S0
2
= Ri+ (zm ) W+ (zm , z0 ) S0 ,
∗
is a Toeplitz matrix, constructed from equation (6.6), given by
where F−
i (zm , z0 )
+
F−
i (zm , z0 ) = ∆W(zm )Wi (zm , z0 ).
(6.29)
Note that convolution result (6.28) represents the focus point response with the correct propagation properties doubled (the error matrix ∆W(zm ) has been removed by
126
-1500
0
6.3 Operator updating
-750
0
750
1500
0.5
1.0
1.0
a. c = 2000 z = 800
-750
0
750
1500
0.5
0
750
1500
1.0
b. c = 2000 z = 640
-1500
0
1.0
d. c = 1800 z = 800
-750
0
750
0.5
1500
-750
0
750
1500
Figure 6.17
0
750
1500
c. c = 2000 z = 960
-1500
0
1.0
e. c = 1800 z = 640
-1500
0
1.0
-750
-750
0
750
1500
0.5
-750
0
750
0.5
g. c = 2200 z = 800
-1500
0
0.5
0.5
-1500
0
1.0
-750
0.5
-1500
0
1.0
-1500
0
1500
f. c = 1800 z = 960
-1500
0
-750
0
750
1500
0.5
h. c = 2200 z = 640
1.0
i. c = 2200 z = 960
2-Dimensional convolution between the CFP gather and its operator for different erroneous operators. Note that all convolutions give the same result, meaning that erroneous CFP gathers can be automatically updated.
the convolution). Hence, if the first focusing step is followed by the convolution process then the WRW response has been modified to a RW2 response, despite the use
of an erroneous synthesis operator. The traveltimes for the correct focusing operator
are obtained from the convolution result by halving the time scale. By making use
of the 2-dimensional convolution the assumption is made that the medium above
the focus point is laterally invariant, which means that W+ (zm , z0 ) and ∆W(zm )
are both Toeplitz matrices (see also the discussion after the derivation of equation
(6.8)).
The 2-dimensional convolution between the CFP gather and its synthesis operator
is shown in figure 6.17 for the same erroneous synthesis operators as in figure 6.3.
The time and offset of the convolution results are displayed by halving the original
spatial and time sampling rate. Note that all convolution results are in traveltime
identical with the correct focusing operator. Note also that the amplitudes in the
different convolution results are weaker for the positive velocity errors. This means
that the Fresnel zone is shifted outside the aperture in the first focusing step.
The operator updating by the 2-dimensional convolution can be interpreted as a
time interpolation in plane wave components. By making use of the linear Radon
transform the erroneous focusing operator and its CFP response can be expressed
in terms of intercept time ∆τ for each ray-parameter component p. The correct
focusing operator can then be found by averaging for each p−value the intercept
Chapter 6: Operator updating
one-way time [s]
-1500
0
-750
lateral position [m]
0
750
1500
0.5
-1500
0
127
-750
1.0
c = 1800 [m/s] z = 640 [m]
c = 1800 [m/s] z = 960 [m]
⇓
τ − p transform
intercept time [s]
-50
-25
ray parameter [ms/m]
0
25
50
0
0.5
1.0
1.0
⇓
-50
-25
first iteration
intercept time [s]
-50
-25
0
0.5
25
⇓
τ − p transform
0.5
0
1500
0.5
1.0
0
lateral position [m]
0
750
ray parameter [ms/m]
0
25
⇓
50
first iteration
50
0
-50
-25
0
25
50
0.5
1.0
1.0
result after first update
Figure 6.18
result after first update
The correct focusing operator can be found by choosing for each p value the intercept
time exactly in the middle of the erroneous operator and erroneous response intercept
times. Note that only one iteration is needed to obtain the correct focusing operator.
times of the time-reversed focusing operator and the related focus point response.
Averaging in the τ − p domain means that the focusing operator and its focus point
response are compared in points with equal slopes. An example for the updating
scheme in τ − p is shown in figure 6.18. For a 1-dimensional medium the correct
operator is obtained in one step.
So there are two methods to update an erroneous focusing operator to a correct
focusing operator: interpolation in offset or interpolation in ray-parameter. The
difference of the two methods is explained in figure 6.19. If the assumption of a
128
6.3 Operator updating
0
time [s]
offset
τ −p
0.5
1.0
-1500
-750
0
750
lateral position [m]
0
0
τ − p interpolation
time [s]
1500
offset interpolation
0.5
1.0
-1500
Figure 6.19
0.5
-750
0
750
lateral position [m]
1500
1.0
-1500
-750
0
750
lateral position [m]
1500
Operator updating by averaging the times for common ray-parameter (left) or for common offset traces (right). The ray-parameter method converges for a 1D medium within
one step to the correct operator. Note the small difference with the offset interpolation
method.
laterally invariant medium is valid then the ray-parameter scheme will converge in
one iteration to the correct operator. The method to obtain the correct operator is
the same for both methods: time convolution of the erroneous operator with the CFP
gather and halving the time scale of the result. This method works in the frequency
domain by multiplication of the CFP gather with its operator. A disadvantage of
the frequency domain implementation is that is takes a lot of computation time to
transform the data and the operator to the frequency domain and the convolution
result back to the time domain. An advantage of the frequency domain method is
that it can handle complex operators.
An alternative and more faster method is based on (automatic) tracking of the erroneous response in the CFP gather. In this way all the computation can be done
in the time domain (see appendix C). However, depending on the complexity of the
medium the time domain method can give problems with complicated operators.
Morton and Thorbecke (1996) have shown that it is possible to do automatic operator updating for 3-dimensional numerical seismic data. A flow scheme of this
operator driven migration process is shown in figure 6.20. This operator driven migration process takes properly into account any unknown traveltime effects due to
ray bending, anisotropy or anything else that may occur in the subsurface. This can
be easily understood if one bears in mind that the focusing operators are directly
determined from the data. Hence, in the CFP method the focusing operators in the
Chapter 6: Operator updating
129
input data
first
focusing
step
operator
updating
CFP gather
move-out
analysis
image gather
second
focusing
step
image
Figure 6.20
Operator driven migration process. The output of the CFP migration is not in depth
but in one-way image time.
migration process are not limited by the restrictions inherent to any user-specified
parametric subsurface model. A subsurface model need only be specified in the
post processing step for time to depth conversion of the one-way time image. This
subsurface model can be obtained by a global inversion of the estimated focusing
operators. The calculation of the global inversion is beyond the scope of this thesis.
6.3.2
Dipping reflector
In the previous section the operator updating was explained with a flat reflector,
in this section the updating for a dipping reflector is discussed. In figure 6.21 the
response of a flat reflector and a dipping reflector are shown. The same model is used
as before; velocity above the reflector 2000 m/s and the reflector depth is 800 m, the
dip is chosen at 20◦ . In both examples an erroneous focusing operator is used with a
velocity of 1600 m/s and a depth of 640 m. In figure 6.21c,d the response in the CFP
gather is shown together with its erroneous operator (the spiky event). Note that
due to the shadow zone the left hand side in the dipping reflector response does not
contain useful information. The focusing operator is updated from the erroneous
CFP gather with the two methods discussed in the previous section. The offset
130
6.3 Operator updating
interpolation shows after one iteration still a deviation from the correct operator
-1500
0
-750
750
lateral position [m]
0
750
1500
-1500
0
750
●
1500
one-way time [s]
-1500
0
1500
●
-750
0
b. dipping reflector model
750
1500
0.5
-1500
0
-750
0
750
1500
0.5
1.0
1.0
c. CFP gather, c = 1600 and z = 640
-1500
0
one-way time [s]
lateral position [m]
0
750
1500
a. flat reflector model
-750
0
750
1500
0.5
d. CFP gather, c = 1600 and z = 640
-1500
0
-750
0
750
1500
0.5
1.0
1.0
e. operator updating in offset
-1500
0
one-way time [s]
-750
-750
0
750
f. operator updating in offset
1500
0.5
-1500
0
-750
0
750
1500
0.5
1.0
1.0
g. operator updating in ray-parameter
Figure 6.21
h. operator updating in ray-parameter
Operator updating for a flat (left) and dipping (right) reflector. Ray-parameter interpolation gives in one iteration the correct operator for the flat reflector. The updated
operator for the dipping reflector seems to deviate from the correct operator.
Chapter 6: Operator updating
one-way time [s]
-1500
0
response
update
operator
131
-750
lateral position [m]
0
750
1500
0.5
1.0
a. updating points
one-way time [s]
-1500
0
-750
b. focus point response
0
750
0.5
1500
-1500
0
-750
0
750
1500
0.5
1.0
1.0
c. comparison same lateral position
Figure 6.22
d. comparison correct lateral position
Updating of a dipping reflector response. Note that the updated operator (in τ − p
domain) has a lateral shift with respect to the original focus point position.
(the spiky event). The ray-parameter interpolation method gives in one iteration
the correct focusing operator for the flat reflector (g). However, the ray-parameter
interpolation method for the dipping reflector still gives a deviation from the correct
operator (h). Careful inspection of the apex of the hyperbola in figure 6.21h shows
that the apex of the updated operator is not positioned at the same lateral position
as the apex of the correct operator.
To show what happened for the updated focusing operator in the dipping reflector
example the same dipping model is used as before but now with only a depth error in
the focusing operator (depth error of + 20 % = 960 m). In figure 6.22a three dots are
drawn in the correct subsurface model, the black dot represents the position of the
erroneous focusing operator, the white dot represents the focus point response and
the gray dot is the position of the updated operator obtained after ray-parameter
interpolation. The position of the erroneous focus point response is, analogous to
the flat layer, on the opposite side of the reflector. Due to the updating in the rayparameter domain, thus interpolation of the intercept-time for waves with the same
ray-parameter value, the updated operator is positioned exactly in the middle of the
two points. This updated operator has, for a layer with a constant dip around the
focus point, its focus point exactly on the reflector. However, the lateral position of
the focus point of the updated focusing operator is shifted down-dip (for a negative
depth error the shift is up-dip) in comparison with the initial focusing operator. In
132
-1000
lateral position [m]
-500
0
500
1000
500 ➊
1000
➋
0
one-way time [s]
0
6.3 Operator updating
➌
one-way time [s]
1000
0.5
b. focusing operator for ➊
a. subsurface model
-1000
-500
0
500
1000
➊
➋
0.5
0
-1000
-500
0
500
1000
lateral position [m]
-500
0
500
1000
➊
➋
0.5
➌
1.0
1.0
c. CFP gather for ➊
0
one-way time [s]
lateral position [m]
-500
0
500
1.0
1500
0
-1000
-1000
-500
0
d. updated operator for ➋
500
1000
0
0.5
➊
➋
500
➌
1000
1.0
-1000
1500
e. CFP gather for ➋
Figure 6.23
f. updated position
Illustration of the update procedure for a simple model. The focusing operator for the
first boundary has been chosen as an initial estimate for the second boundary.
figure 6.22b the focus point response is shown together with the focusing operator.
The updated operator after the ray-parameter interpolation step is shown in figure
6.22c, the spiky event represents the operator at the same lateral position as the
initial operator. Note that the updated operator has a lateral shift with respect
to the original focus point position. The lateral shift can be calculated by using
the interpretation shown in figure 6.22a. Taking this shift into account the shifted
focusing operator on the reflector is identical with the updated operator which is
shown in figure 6.22d.
A nice example of the operator updating scheme is shown in figure 6.23. In this
example the updated focusing operator for the first depth level is used as an estimate
for the next one. The example shows that in CFP migration a CFP gather should
not be computed only along its time-reversed synthesis operator but it should be
Chapter 6: Operator updating
133
computed around its time-reversed focusing operator in order to avoid loosing the
focus point response (Berkhout, 1996b) and to be able to jump to the next interface.
Once the updated focusing operator is computed the second focusing step can be
carried out. In the imaging step, which was discussed in section 5.3, the result of the
second focusing step is positioned at the lateral position of the focus point. However,
due to the updating in the first focusing step the lateral position of the focus point
can be shifted in the case of a dipping reflector. The lateral shift of the focus point
is dependent on the dip of the reflector. There are two ways to compensate for this
lateral error in the final image; in the global inversion of all the focusing operators
to a (depth) macro model the local dip of the reflector can be computed from the
one-way time image and be taken into account in the global inversion. Another way
is to provide an estimate of the local dip during the imaging, and taking the correct
position of the updated focus point into account in the positioning in the one-way
time image.
By making use of the lateral shift in the focus point response of an erroneous focusing
operator it is possible to estimate the local dip of the reflector. In figure 6.24 the dip
estimation procedure is explained. Two focus points are chosen above (or below) the
reflector of interest at the same depth level but with a different lateral position. The
focus point response is calculated for both focus points. Assuming that the medium
around the area of the two focus points is laterally invariant in the direction of the dip
then the local dip of the reflector can be estimated by using the 2-dimensional crosscorrelation between the two CFP gathers. The CFP gather can be interpreted as the
response of a point below the reflector, where the focus point (the black dot) and the
focus point response (the white dot) are positioned on a line perpendicular on the dip
of the reflector. By using the lateral distance between the two focus points and the
difference in arrival time of the focus point responses the local ray-parameter [s/m]
value of the reflector can be determined. The lateral distance between the two focus
points is chosen by the user, the difference in arrival time between the two focus point
responses is measured from the 2-dimensional cross-correlation. If an estimation of
the velocity is available the reflector angle in degrees can be computed. In figure
6.24 the CFP gathers are computed with erroneous focusing operators (defined with
a velocity −20% = 1600 m/s and at depth level 640 m) which are 150 m separated
from each other. From the 2-dimensional cross-correlation shown in figure 6.24d
the time delay ∆t = 0.048 is determined. With this time delay the angle can be
computed by using the correct velocity of the medium above the reflector resulting in
the correct angle of 20◦ . The sensitivity of the method dependents on the accuracy
of the arrival time measurement from the cross-correlation, an error of one time
sample (4 ms) gives a deviation of 2◦ = 10% in the example of figure 6.24. Note
that the erroneous macro model used to generate the focusing operators does not
distort the dip estimation in the example.
134
6.4 Second focusing step
∆x
φ
2φ
c∆t
∆x
a. principle of dip estimation
one-way time [s]
-1500
0
-750
lateral position [m]
0
750
1500
0.5
lateral position [m]
0
750
1500
1.0
c. CFP gather for operator at x = −75 [m]
one-way time [s]
-750
0.5
1.0
-0.2
-1500
0
-400
lateral position [m]
-200
0
200
d. CFP gather for operator at x = +75 [m]
400
∆x = 150 [m]
-0.1
∆t = 0.048 [s]
0
sin (2φ) = pc =
p = 3.2e
0.1
−4
c∆t
∆x
with c = 2000, φ = 20◦
0.2
e. 2 D cross-correlation of c ∗ d
Figure 6.24
6.4
Local reflector dip estimation with two focus points at the same depth level. Note that
the principle works with any pair of (erroneous) focusing operators defined at the same
depth level.
Second focusing step
The influence of erroneous focusing operators in the second focusing step is of importance if the updating step is not used and one is interested in the quality of the
resulting one-way time image. In the previous chapter the property was observed
that all events in a one-way time image are imaged at the correct one-way image
time independent of the used operators. To examine this property one trace out of a
CFP gather is calculated for one shot record for different depth and velocity errors
in the focusing operator. The same single reflector model is used as before . In figure
6.25b the used shot record, with the source xs = 0 and the receivers positioned in
[−1500, 1500] with a receiver distance of 15 m, is shown. The one-way image time
for the reflector is at 0.4 s, this time is found in figure 6.25b at the trace belonging
Chapter 6: Operator updating
-750
800 [m]
lateral position [m]
0
750
1500
-1500
0
two-way time [s]
750
●
1500
lateral position [m]
0
750
1500
2
200
0
one-way time [s]
-750
1
a. subsurface model
400
focus point depth [m]
600
800
1000
b. shot record xs = 0
1200
1400
200
0
400
focus point depth [m]
600
800
1000
1200
1400
1200
1400
0.5
0.5
1.0
1.0
c. CFP trace xs = 0
200
0
one-way time [s]
depth [m]
-1500
0
135
400
focus point depth [m]
600
800
1000
d. one-way imaging of c
1200
1400
200
0
400
focus point depth [m]
600
800
1000
0.5
0.5
1.0
1.0
e. CFP trace xs = 1500
Figure 6.25
f. one-way imaging of e
The influence of depth errors in the imaging step. Note that for xs = 0 (d) all move-out
corrected traces are positioned at the correct one-way image time independent of the
depth of the focus point.
to the correct depth of 800 m. Every trace in figure 6.25c represents a CFP trace
calculated with a focusing operator for a certain depth level. The CFP traces calculated for the shot with xs = 0 represent the top of the apex of the erroneous focus
point response. Figure 6.25d represents the traces in (c) after move-out correction
with the erroneous focusing operator followed by positioning at one-way image time
(again defined by the erroneous focusing operators). Note that all traces are now
positioned at the correct one-way image time independent of the depth of the focus
point. The CFP trace calculated for an other source position (xs = 1500) shown in
figure 6.25f does not align up at the correct one-way image time.
The observed property can be explained by considering only the stationary phase
contribution in the first focusing step. For the focus point in the middle of the flat
6.4 Second focusing step
750
-750
lateral position [m]
0
750
1500
●
-1500
0
lateral position [m]
0
750
1500
1
a. subsurface model
1200
0
one-way time [s]
-750
2
1500
focus point velocity [m/s]
1600
2000
2400
b. shot record xs = 0
2800
1200
0
focus point velocity [m/s]
1600
2000
2400
2800
0.5
0.5
1.0
1.0
c. CFP trace xs = 0
1200
0
one-way time [s]
depth [m]
-1500
0
two-way time [s]
136
focus point velocity [m/s]
1600
2000
2400
d. one-way imaging of c
2800
1200
0
focus point velocity [m/s]
1600
2000
2400
2800
0.5
0.5
1.0
1.0
e. CFP trace xs = 1500
Figure 6.26
f. one-way imaging of e
The influence of velocity errors in the imaging step. Note that for xs = 0 (d) all moveout corrected traces are positioned at the correct one-way image time independent of
the velocity above the focus point.
reflector the stationary phase trace for the shot record at xs = 0 is xr = xs = 0.
This trace (Td ) is corrected with the time defined by the focusing operator (Td − Ts ).
For the imaging step the trace after the first focusing step is corrected again with the
time defined by the focusing operator (= Td −2∗Ts), positioned at the one-way image
time (Ts ) and renumbered to one-way times (Ti = 0.5∗(Td −2∗Ts )+Ts = 0.5∗Td. So
imaging with an erroneous focusing operator places the stationary phase contribution
of the shot record (for the focus point of interest) at the correct one-way image
time. Note, that this procedure can also be carried out by a summation over all
the (uncorrected) traces in the shot gather followed by halving the time scale. This
procedure selects the stationary trace for a focus point defined at the source position.
By using more accurate focus points the contributions of the different shot gathers
Chapter 6: Operator updating
137
in the CFP image gather will be better aligned and the image quality will be better.
Note also that due to an erroneous focusing operator the Fresnel zone can be shifted
outside the aperture which will give artifacts in the one-way time image.
The same experiment as shown in figure 6.25 is repeated for a dipping layer and
velocity errors in the focusing operator. Again the stationary phase trace originates
from the shot record with xs = 0. The imaged traces for this shot record are shown
in figure 6.26. Note that for positive velocity errors (c > 2000 m/s) the Fresnel zone
around the stationary phase position is quickly shifted outside the aperture and the
strongest contribution in the CFP trace is an artifact. The contribution from the
shot record with its source position at xs = 1500 shows only the correct trace for the
correct velocity. If the Fresnel zone remains within the aperture the velocity error in
the focusing operator has no influence on the one-way image time of the stationary
phase trace.
The foregoing analysis has shown that the time position in the one-way time image
is not very sensitive for errors in the focusing operator. However, the image quality
is very sensitive for erroneous operators. This sensitivity of the CFP image gather
is used in the coherency measurement discussed before.
138
6.4 Second focusing step
Chapter 7
Numerical data examples
To give more insight in the CFP processing techniques several numerical data examples are used. Section 7.1 shows the focus point response for a single diffraction
point in the subsurface for several erroneous focusing operators. A 1-dimensional
medium based on a well log measurement is shown in section 7.2. With this model
the construction of a CFP image gather is illustrated. In section 7.3 a model is
chosen which defines a multi-valued focusing operator for a deep reflector. This
example is used to show the limitations of a simple implementation of the operator
updating procedure. The influence of groundroll and a weathered layer on the CFP
gather is shown in section 7.4. In section 7.5 the syncline model is used to show
the integration of CFP gathers and pseudo VSP data. The integration is shown to
give more insight in the illumination aspects of the CFP gather. In the last section
the Marmousi and the SEG/EAGE salt dome model are used to compare the CFP
image gather with imaging results obtained with other imaging methods.
7.1
Diffraction point
The diffraction point is a suitable experiment for testing the construction and imaging of a CFP gather. In a homogeneous medium a diffraction point is chosen in
the middle of the model at a depth of 500 m and a fixed acquisition spread at the
surface is used. In figure 7.1 several CFP gathers are shown which are constructed
with different synthesis operators. Note that only the focusing operator with its
focus point defined at the diffraction point gives the correct focus point response.
For a positive (negative) depth error in the synthesis operator the constructed event
in the CFP gather is positioned above (below) the correct position. This is shown
in figure 7.1g and e, where the black line represents the operator and the event with
the wavelet represents the focus point response. In the updating procedure, where
the updated operator is chosen in between the operator and the CFP gather, the
correct operator will be obtained in one step.
140
7.1 Diffraction point
lateral position [m]
0
750
1500
-1500
0
two-way time [s]
-750
✮
500
-750
lateral position [m]
0
750
1500
0.5
1.0
1000
a. subsurface model
one-way time [s]
-1500
0
-750
0
b. shot record with source at x = 0
750
1500
0.5
-1500
0
0
750
1500
1.0
c. correct focus point response
-1500
0
one-way time [s]
-750
0.5
1.0
-750
0
750
d. focus point response for x = 100, z = 500
1500
0.5
-1500
0
-750
0
750
1500
0.5
1.0
1.0
e. CFP with focus point at x = 0, z = 300
-1500
0
one-way time [s]
depth [m]
-1500
0
-750
0
750
1500
0.5
f. focus point response for x = 200, z = 500
-1500
0
-750
0
750
1500
0.5
1.0
1.0
g. focus point response for x = 0, z = 700
Figure 7.1
h. focus point response for x = 300, z = 500
CFP gather for a single diffraction point in the subsurface constructed with different
synthesis operators. The scaling in the pictures in the CFP gathers for the depth errors
(0.1) and the lateral errors (0.02) is not the same. The black line indicates the focusing
operator. Note the two events in the CFP gather for a focusing operator with a lateral
error.
Chapter 7: Numerical data examples
two-way time [s]
-1500
0
-750
lateral position [m]
0
750
1500
-1500
0
0.5
0.5
1.0
1.0
a. synthesis for shot at x = −750
time [s]
-1500
-0.5
-750
0
750
1500
-1500
-0.5
0
0.5
0.5
Figure 7.2
lateral position [m]
0
750
1500
b. synthesis for shot at x = 0
0
c. bifocal result of figure 7.1d
-750
141
-750
0
750
1500
d. bifocal result of figure 7.1h
The CFP corrected result between the operator with a lateral shift of 300 [m] and a shot
record with its shot position at x = −750 (a) and at x = 0 (b). In the bifocal results (c)
and (d) part of the imaging finite aperture artifacts can be recognized.
For a synthesis operator with a lateral error the CFP gather is shown in figure 7.1d,
f and h for a lateral error of respectively 100, 200 and 300 m. In the updating procedure the operator can converge, after a few iterations and selecting consequently
the same event, to the correct synthesis operator, but in the second focusing step
step the focus point response will be positioned at the wrong lateral position, namely
the position of the initial operator. This position error can be taken care of if one
determines the shift of the apex of the updated focusing operator with respect to
the initial focusing operator. Note, that the amplitudes of the focus point responses
with a lateral error have a maximum amplitude 5 times lower than the responses
with a depth error. This lower amplitude is an indication that the energy of the
focus point response is low for an operator not positioned at the diffraction point
position. To investigate the origin of the two events visible in the focus point responses for a focusing operator with a lateral error the CFP corrected shot records
are investigated.
In figure 7.2a the synthesis result between the operator with a lateral shift of 300
m and a shot record with its shot position at x = −750 and a shot position at
x = 0 is shown in figure 7.2. In these figures we see two zones on the edges of the
acquisition aperture which give the double contribution in the CFP gather. For the
ideal situation the integration of the results shown in figure 7.2a and b should give
no contribution in the CFP gather. If the aperture would be extended to infinity
142
7.2 One dimensional multi layer model
the result at the edges will asymptotically go to a Fresnel zone, but never reach it
The bifocal result (2D cross-correlation between operator and CFP gather) for the
CFP gather are shown in figure 7.2c and d. From these bifocal results it is more
clearly observed that the results obtained with a shifted focusing operator represent
the artifact for the correct image of the diffraction point. From this last observation
it is concluded that the double events present in the CFP gathers shown in figure
7.2 originate from the finite acquisition aperture.
7.2
One dimensional multi layer model
The shot record shown in figure 7.3a is modeled in a 1-dimensional medium based
on a well log shown in figure 7.3b and is used as input data for the CFP processing
scheme. Due to the presence of the many layers in the model this synthetic data set
is used to clarify the concept of CFP image-gathers. Choosing a synthesis operator
at the top (➊), middle (➋) and bottom (➌) of the model and scaling the constructed
CFP gather to one-way times gives the results shown in figure 7.3d, e and f. The
scaling is carried out with respect to one-way image time; the traveltime between
the focus point and the source position at the surface.
In figure 7.3d, e and f it is observed that in the focusing area of the synthesis
operator the CFP-gather gives a good representation of the subsurface. Further
away from the focusing area the construction is less accurate and finite aperture
artifacts distort the events in the CFP gather. Constructing the CFP image-gather
from these CFP gathers is carried out by a time selection and time correction, given
by interpolated focusing operators, for all time samples. This construction is shown
in steps in figure 7.3g, h and i. In figure 7.3g the three single focus gathers d, e and
f are combined to one multiple focus gather. This is done by combining the samples
around a window of the focusing operator times and placing these selections into one
file. The overlap between the contributions of different CFP gathers is chosen at a
few time samples. The following step, from figure 7.3g to h, is a move out correction
defined by the focusing operators used in the first focusing step. In figure 7.3h it is
observed that only the events at the operator times are perfectly flat and will give
a good representation of the reflector present at the defined focus point. The events
in between the focus points of the first focusing step are not very well aligned. Thus
for the second focusing step a focusing operator is needed at every time sample.
These focusing operators are obtained by linear interpolation, for every offset, of
the operators from the first focusing step. The result of this second focusing step is
shown in figure 7.3i. In this gather the events around the focus point are all aligned,
the events in between focus points are only good aligned for the deeper events. The
operators which have a focus point at deeper layers give a better alignment due to
the smaller move out correction for the higher offsets. The shallow part shows a
significant wavelet stretching for the higher offsets due to the convergence of the
Chapter 7: Numerical data examples
lateral position [m]
800
1600
0
←➊
➊
1000
depth [m]
1
two-way time [s]
lateral position [m]
-1600
0
1600
0
0
2
1000
←➋
2000
3
3000
4
4000
depth [m]
0
0
143
➋
2000
3000
←➌
0
a. shot record
0
800
4000
b. velocity log
1600
0
0
800
➌
1600
0
c. subsurface model
0
800
1600
one-way time [s]
➊
1
➋
1
1
➌
2
one-way time [s]
0
2
d. focus at ➊
0
800
1600
1
2
0
2
e. focus at ➋
0
800
1600
1
g. combination of ➊+➋+➌
Figure 7.3
2
0
f. focus at ➌
0
800
1600
1
h. move out correction of g
2
i. CFP image-gather
Construction of CFP image-gathers in a 1-dimensional medium. The arrow indicates
where the focus points are chosen. It is interesting to see that only 3 focus points yield
already a reasonable CFP image-gather. Note the stretching of the wavelet due to the
convergence of the traveltime curves at the higher offsets.
0
0
800
1600
d. ∆T = 128 ms.
Figure 7.4
0
800
1600
1
2
a. ’perfect’ response.
1
2
0
0
0
800
1600
e. ∆T = 256 ms.
0
800
1600
1
2
b. ∆T = 32 ms.
1
2
0
one-way time [s]
1600
one-way time [s]
800
1
2
one-way time [s]
0
one-way time [s]
one-way time [s]
0
7.2 One dimensional multi layer model
0
one-way time [s]
144
c. ∆T = 64 ms.
0
800
1600
1
2
f. ∆T = 512 ms.
CFP image gathers in a 1-dimensional medium for different sampling rates, in one-way
image time, of the focus points. Note that the CFP image-gathers clearly identify that
the focusing operators around 1 second are not correct.
time curves of the synthesis operators at higher offsets. These stretched events are
normally not included in the construction of the image and are muted out for stretch
values greater than a certain threshold. To have a better alignment for the shallow
part more focus point must be used to construct the CFP image-gather.
The influence of the number of operators, used in the first focusing step, on the quality of the CFP image-gather is shown in figure 7.4 for several numbers of operators.
In figure 7.4a the time response of the ideal CFP image-gather is shown (which is for
a 1-dimensional medium equal to the zero offset trace displayed in one-way time).
By constructing CFP image-gathers with different sampling rates, in one-way image
time, for the focusing operators in the first focusing step and comparing these results
with the ideal gather (a) gives an indication of the focusing area of the synthesis
operators. Comparing the result of figure 7.3i with the double focus gather in figure
Chapter 7: Numerical data examples
0
0
lateral position [m]
800
1600
0
0
145
lateral position [m]
800
1600
0
0
lateral position [m]
800
1600
2
one-way time [s]
one-way time [s]
one-way time [s]
1
1
1
3
4
a. CFP gather z = 2000 [m]
Figure 7.5
2
b. updated operators
2
c. operator collection
Operator updating in a 1-dimensional medium is very easy; within one step all events
can be updated to the correct focusing operator.
7.4a shows already that three synthesis operators are not sufficient to illuminate
the whole depth range. However, for the focus points defined in the deeper parts
of the model the focusing area is bigger than for a focus point close to the surface.
By selecting a focus at every 32 ms (one-way time) in the subsurface a better CFP
image-gather is built up which is shown in figure 7.4b. Note that in this gather the
stretched events are muted out of the result. To connect the contributions from the
different CFP gathers an overlapping window with a cosine taper is used. Comparing the results in figure 7.4 indicates that a focus point sampling of 64 or 128
ms is sufficient to illuminate all events in between the defined operators. Note also
that the CFP image-gathers clearly identify that the focusing operators just before
1 second are not correct.
Operator updating in a 1-dimensional medium is very easy; from one (erroneous)
CFP gather all other focusing operators could be determined. This is due to the
fact that for a 1-dimensional medium the operator updating procedure in the τ − p
domain always converges in one step to the correct operator for all events in the CFP
gather (see the discussion after equation (6.28)). Figure 7.5a shows a CFP gather
constructed with a focusing operator defined at z = 2000 m. After 2-dimensional
convolution between operator and CFP gather and a renumbering to one-way time
and CFP offset gives the collection of updated operators shown in (b). Note that due
to the erroneous operator for the levels above the correct operator the Fresnel zone
is shifted outside the acquisition aperture. For comparison a collection of correct
traveltime curves is shown in (c). Note that this updating method can be used for
an initial guess of the focusing operators in a more complicated medium.
A reflection coefficient gather of the 1-dimensional model can be constructed according to the scheme given in figure 7.6. To show the principle of the construction of
146
7.2 One dimensional multi layer model
shot records
synthesis
operators
first
focusing
step
CFP gather
crosscorrelation
bifocal image
t=0
selection
reflectivity gather
τ-p
transform
reflection coefficient gather
Figure 7.6
Scheme for the construction of a reflection coefficient gather. The selection around t =
0 consists of a small time window defined by the one-way time distance between the
operators.
the reflection coefficient gather only the density well log measurement is used while
the velocity is chosen constant at c = 2000 m/s. At every 8 m (= 4 ms) in depth
of the subsurface a focus point is chosen and a focusing operator is calculated. The
operator is used to construct the CFP gather for the defined focus point. The construction of the CFP gather is followed by a 2-dimensional cross-correlation between
the operator and the CFP gather. From the cross-correlation result the t = 0 level is
selected and placed in the reflectivity gather. This process is repeated until all focus
points have been handled. The resulting collection of t = 0 selections is shown in figure 7.7a and is called the reflectivity gather. This reflectivity gather contains all the
reflection information for the same lateral position in the model. A transformation
of the reflectivity gather to the τ − p domain gives the reflection coefficient gather
Chapter 7: Numerical data examples
0
-300
-150
position [m]
0
150
300
0
intercept time [s]
time [s]
ray-parameter [ms/m]
-25
0
25
50
0.5
0.5
1.0
1.0
1.5
1.5
2.0
-50
147
2.0
a. reflectivity gather
Figure 7.7
b. reflection coefficient gather
The reflectivity gather, for a 1D model based on density contrasts only, constructed according to the scheme of figure 7.6. The reflection coefficient gather (b) is obtained by a
τ − p transformation of (a).
which is shown in figure 7.7b. This procedure is similar to the one discussed in
section 5.4.2. Note that the procedure to obtain the reflection coefficient gather can
be made more efficient by selecting more samples from the cross-correlation result.
The reflection coefficient gather is the input for the lithologic inversion scheme as
shown in the introduction.
7.3
’Void’ model
The model shown in figure 7.8a is developed to test the operator updating procedure
which was introduced in the previous chapter. A dipping reflector is positioned below
a velocity ’void’ which distort the reflection in a such a way that the reflected event at
the surface contains multiple arrivals. The dipping reflector is a line defined through
the points (x = −1500, z = 600) and (x = 1500, z = 975), where the velocity above
the reflector is 2000 m/s, below the reflector 2200 m/s and in the middle of the ’void’
the velocity is 1000 m/s. The shot record with a source position at x = 0 is shown
in figure 7.8b. Note the breaking of the reflection event originating from the middle
part of the dipping reflector. The two events above the main reflection event are due
to the breaking in three parts of the downgoing source wave field. The middle part
of the downgoing wave field gives rise to the reflection observed in the middle of the
shot record. The two other downgoing wave fields give rise to the events above and
aside of this main reflection event.
A focus point is chosen on the reflector at x = 0 and two methods to calculate a CFP
gather are used; one in the frequency-domain and one in the time-domain, where the
last method only takes the first arrival times into account (see Appendix C about
148
7.3 ’Void’ model
-1500
0
-750
lateral position [m]
0
750
1500
two-way time [s]
depth [m]
c=2000
← c=1000
500
-1500
0
c=2200
-750
lateral position [m]
0
750
1500
1
2
a. subsurface model
one-way time [s]
-1500
0
-750
0
b. shot record with xs = 0
750
1500
0.5
-1500
0
0
750
1500
0.5
1.0
1.0
c. focusing operator
-1500
0
one-way time [s]
-750
-750
0
0.5
d. CFP gather frequency implementation
750
1500
-1500
0
-750
0
750
1500
0.5
1.0
1.0
e. focusing operator
Figure 7.8
f. CFP gather time implementation
A subsurface model with a velocity ’void’ distorts all reflection events below the ’void’.
Note that if the correct model is used there is no big difference in the CFP gather between
the frequency and time method.
the different implementations of the CFP method). The focusing operator which is
used in the frequency method is shown in figure 7.8c and the operator used in the
time-domain scheme is shown in figure 7.8e. Note that the traveltime operator is
only defined at the first arrival times. The CFP gather obtained with the frequency
method (d) and the time method (f) are, with respect to the first arrival times, not
very different. This is an important observation because it means that, at least for
this model, first arrival times are sufficient to calculate the focus point response.
For the later arrival times the frequency method gives a focus point response which
resembles the events in the focusing operator.
In the next example an erroneous initial focusing operator is used to investigate
whether the operator updating method from the previous chapter can be used to
Chapter 7: Numerical data examples
one-way time [s]
-1500
0
-750
0
750
1500
-1500
0
750
1500
1.0
1.0
a. erroneous CFP gather
-1500
0
one-way time [s]
0
0.5
0.5
-750
0
b. an attempt to update (a)
750
0.3
0.6
-750
149
1500
-1500
0
-750
0
750
1500
0.3
c. image with erroneous operators
Figure 7.9
0.6
d. image with ’correct’ operators
CFP gather constructed with an erroneous focusing operator. Note that updating of the
operators will not lead to the correct operators and therefore an incorrect image will be
obtained (c).
update the (first) arrival times of the operator. The erroneous operator is calculated
for a homogeneous medium with a velocity of 2000 m/s. In the resulting CFP gather,
shown in figure 7.9a, it is observed that updating of the operator in the middle part
of the model is complicated. In an attempt of tracking the focus point response
an error is easily made, therefore the automatic operator updating, by means of
a 2-dimensional convolution as explained in section 6.3, is prefered. The updated
CFP gather is shown in figure 7.9b and shows only the correct operator times at
the sides of the model, the middle of the operator is still not good defined. In the
foregoing example it has been shown that for a complicated subsurface model and
a complicated focusing operator it is difficult to obtain a correct operator from an
erroneous CFP gather. In these difficult situations the user must be able to define,
from the erroneous CFP gather, the updated operator times from an interactive
display environment. The influence of an erroneous operator on the CFP image
quality is shown in figure 7.9c, where a homogeneous model is used for the calculation
of the operators defined on the dipping reflector. The image obtained with the
’correct’ (only first arrival times) operators is shown in figure 7.9d. From these
two CFP images it is concluded that for the defined example the use of erroneous
focusing operators is disturbing the quality of the CFP image significantly.
150
7.4 Weathered layer model
-750
lateral position [m]
0
750
1500
↑
c=750
c=2000
250
500
c=2500
one-way time [s]
-750
lateral position [m]
0
750
1500
0.5
1.0
a. subsurface model
-1500
0
-750
0
b. shot record with xs = 0
750
0.5
1500
-1500
0
-750
0
750
1500
0.5
1.0
1.0
c. CFP gather for (x = 0, z = 100)
Figure 7.10
7.4
-1500
0
two-way time [s]
depth [m]
-1500
0
d. CFP gather for (x = 0, z = 400)
’Removal’ of surface waves (direct wave and dispersive groundroll) by the synthesis
process for focus points below the weathered layer. Note the irregular shaped ’hyperbola’ of the deeper flat reflector.
Weathered layer model
The linear events in a seismic shot record, the direct wave and the (dispersive)
groundroll, are the low velocity events which are not desired in the CFP gather. To
remove these events in a proper way a macro model of the weathered layer should
be estimated. It has been shown that by using the areal shot record technology it
is possible to estimate an effective weathered layer model which accounts for the
propagation properties in the weathered layer (see Thorbecke et al., 1992; Thorbecke, 1994). However, after the synthesis process the linear events have only a small
contribution in the relevant part of the CFP-gather due to their low apparent velocities. This is illustrated in figure 7.10. In this example two different focusing
operators are used which have their focus points at respectively 100 and 400 meter
below the surface. Note that close to the surface of the model a layer with a vertical
velocity gradient is used to make the surface waves more dispersive and simulate
a weathered layer. A fixed acquisition spread is used with a source and receiver
sampling of 10 m and positions at the surface. In figure 7.10c the linear events only
have a contribution in the CFP gather close to zero time. If the focus point is chosen
deeper, as shown in (d), the effects of the direct wave and the groundroll are shifted
to ’negative’ times. In this way the linear near surface effects are not disturbing the
focus point response.
Chapter 7: Numerical data examples
one-way time [s]
-1500
0
-750
0
750
1500
0.5
-1500
0
0
750
1500
0
750
1500
0.5
1.0
1.0
a. erroneous operator
-1500
0
one-way time [s]
-750
151
-750
0.5
0
b. response of (a)
750
1500
-1500
0
-750
0.5
1.0
1.0
c. correct operator
Figure 7.11
d. response of (c)
By using the information in the erroneous CFP gather information of the weathered
layer disturbances can be obtained. Note that the correct focusing operator contains
all the weathered layer disturbances at the receiver side.
The response of an erroneous focusing operator, defined in a model without the
effects of the weathered layer, can be helpful in determining the influence of the
weathered layer. Comparing the (erroneous) operator and the CFP gather gives information about the propagation disturbances due to the weathered layer (Thorbecke, 1992). An iterative updating procedure of the focusing operator can take all
the geometrical propagation effects into account. If the matching of the focusing
operator with the selected event (the flat reflector at 400 m depth) is correct then
the operator of the deeper layer should come out correctly including the disturbing
effects of the weathered layer. The same model as shown in figure 7.10a is used to
illustrate the proposed method.
The initial synthesis operator is a pulse response of a source at 400 m depth in a
homogeneous medium with a velocity of 2000 m/s as given in figure 7.11a. The
CFP gather constructed with this synthesis operator is shown in figure 7.11b. Note
the mismatch between the synthesis operator and the CFP gather. This mismatch
can be used to iterate to the correct operator. After several iterations the correct
synthesis operator shown in figure 7.11c should be obtained, the focus point response
is in travel-time identical with the synthesis operator so the second focusing step can
be carried out.
In the seismic industry sometimes buried sources are used to avoid the disturbances
of the weathered layer at the the source side. In this technique the sources are
152
7.4 Weathered layer model
one-way time [s]
-1500
0
-750
0
750
1500
0.5
-1500
0
0
750
1500
0
750
1500
0.5
1.0
1.0
a. operator for focusing in emission
-1500
0
one-way time [s]
-750
-750
0
750
0.5
b. response of (a)
1500
-1500
0
-750
0.5
1.0
1.0
c. operator for focusing in detection
Figure 7.12
d. response of (c)
By making use of buried sources the weathered layer influence at the source side is
avoided. Note that the difference in depth level between source and detector is clearly
visible between the CFP gathers for focusing in emission (b) and focusing in detection
(d).
positioned below the weathered layer and only the measured wave field at the surface
contains the disturbing influence of the weathered layer. This technique is simulated
for the weathered layer model from figure 7.10, where the sources are positioned
below the weathered layer (at z = 110 m, which is useful for the defined model but
is an unrealistic depth position). The results for focusing in detection are shown in
figure 7.12b and for focusing in emission in figure 7.12d. From these results it can
clearly be seen that the operator for focusing in emission is in traveltime equal with
the focus point response for focusing in detection (and visa versa).
one-way time [s]
-1500
0
-750
0.5
0
-1500
0
-750
0
750
1500
0.5
1.0
1.0
a. focusing in detection
Figure 7.13
b. focusing in emission
By using a random source depth a very rough weathered layer is simulated. Note the
great difference between focusing in emission and focusing in detection.
Chapter 7: Numerical data examples
153
input data
focusing in
detection
focusing in
emission
determine
source statics
determine
receiver statics
surface consistent
statics
further CFP processing
Figure 7.14
Weathered layer correction by making use of the difference in information content
between focusing in emission and focusing in detection.
Note that if the weathered layer disturbances are even stronger as used in the foregoing example then the construction of the CFP gather will fail due to the destruction
of the Fresnel zone. This is shown in figure 7.13 where a ’random’ depth shift, a
maximum depth of 250 m is used to show the effects more pronounced, is applied to
the source positions for shot records containing the response of one single reflector.
Note that for focusing in emission the Fresnel zone is totally destroyed and the individual contributions of the different common receiver gather is visible in the CFP
gather. The CFP gather for focusing in detection show the different static shifts
caused by the changing the depth level of the sources. Note that with the aid of
this gather is is very easy to obtain the (relative) static shift at the source side.
The (relative) static shift for the detectors can be obtained from the CFP gather
for focusing in emission result after the source shifts have been determined from the
CFP gather for focusing in detection. This idea is shown schematically in figure
7.14.
A wave emitted by a source at the surface propagates down, gets reflected from
deeper interfaces and propagates back to the receiver at the surface. It thus travels
at least two times through the weathered layer; the first time when it is emitted
by the source and the second time just before it is recorded at the surface. These
154
7.5 Syncline model
propagation effects through the weathered layer are of a different kind, after emission at the source position the wave travels to a reflector in the far field (related
to the weathered layer) in which the wave field has partly recovered from the disturbed influence of the weathered layer. After reflection, the reflected wave field
travels upward and close to the surface it has to propagate through the weathered
layer again before it can be measured by the receivers. After propagation through
the weathered layer the wave field is measured immediately by the receivers, so the
receivers measure the near field of the weathered layer influence. These assumptions about the propagation behavior of the wave field through the weathered layer
explains why the scheme of figure 7.14 can work (Thorbecke and Berkhout, 1993).
7.5
Syncline model
The syncline model was used in the previous chapter to illustrate the applications
of CFP technology. In this chapter the syncline model is used to illustrate the
connection with other kinds of illumination techniques. The Vertical Seismic Profile
(VSP) is often used to get a more detailed view of the subsurface around the borehole.
By using CFP gathers for focusing in detection, for focus points chosen at different
depth levels in the subsurface and selecting only one lateral position out of these CFP
gathers, a gather can be built up which can be interpreted as a VSP gather. The
-750
lateral position [m]
0
750
1500
500
-1500
0
time [s]
depth [m]
-1500
0
1000
-750
lateral position [m]
0
750
1500
0.5
1.0
∇
1500
a. subsurface model
b. shot gather
c. pseudo VSP
-1500
0
depth [m]
0
500
-750
0
750
1500
0.5
1000
1.0
1500
0
0.5
1.0
time [s]
Figure 7.15
0
0.5
1.0
time [s]
d. CFP gather
One shot record out of the syncline model is used to generate the pseudo VSP (c). The
CFP in (d) has its focus point defined at the indicated position in the model (a).
Chapter 7: Numerical data examples
0
shot record
pseudo VSP
155
CFP gather
time [s]
0.5
➝
1.0
Figure 7.16
The integration between a shot record at the surface (left part) and a CFP gather (right
part) via the pseudo VSP (middle part), for a focus point in the subsurface.
interpretation and use of pseudo VSP data, which is generated from surface data, in
the imaging process can be found in the work of Alá’i (1997). Here, the integration
with the pseudo VSP gather is shown to give more insight in the illumination aspects
of the CFP gather. Figure 7.15 shows the model (a), the shot record (b) used to
generate the pseudo VSP (c) and the CFP with its focus point at the indicated
position in the model. The right picture in figure 7.15c shows the pseudo VSP after
removal of the non-causal events.
Figure 7.16 show the integration between a shot record in the middle of the model
and a CFP gather with a focus point at the position x = 495 m, z = 1200 m in
the model. The left part of figure 7.16 shows the shot record. At the position
of the last receiver, xr = 495 m, of the shot record a pseudo VSP measurement
is simulated by selecting the wavefield at the same lateral position at successively
deeper depth levels obtained with an inverse extrapolation algorithm in the spacefrequency domain. The last part of the figure shows the CFP gather which starts
with the trace at the lateral position of the shot record (xs = 0 m). The pseudo
VSP generation shows that only one branch of the triplication in the shot record,
originating from the syncline, is propagated to the depth position of the focus point.
The other branch, indicated with an arrow, is slowly loosing its energy for deeper
depth levels. The event in the pseudo VSP, which travels from top-left to middleright represents the propagation of the direct source wavefield to the focus point.
The downgoing source wavefield corresponds to the selection of one offset from the
focusing operator per depth level. All events above this line represent non-causal
wavefields, note that these wavefields are also observed in the CFP gather.
156
7.6 Comparison of imaging results for realistic numerical data
7.6
Comparison of imaging results for realistic numerical data
There are several numerical data sets available which are modeled for the purpose of
testing the performance of migration algorithms in a controlled way. The Marmousi
data set is probably the most well-known data set for that purpose. Recently new
modeling results, based on a realistic 3-dimensional velocity model, became available
which can be used to test the imaging result beneath a salt dome. Both numerical
data sets are used in this thesis to show the strength and limitations of the CFP
migration method.
7.6.1
Marmousi model
The Marmousi data set of the Institut Français du Pétrole (IFP, Versteeg and Grau
(1991)) is generated by using a very complicated geologic model, shown in figure 7.17,
and provides a challenge for any migration method, even when the correct velocities
are used. The Marmousi dataset was initially set up as a blind test experiment
to evaluate the quality of the industries inversion and pre-stack depth migration
methods. The data was modeled with a second order (time and space) acoustic
finite difference scheme. The acquisition geometry is a marine type of acquisition
(end of spread) containing 96 geophone groups, with an minimum offset of 200 m
and a receiver spacing of 25 m. The modeled data contained 240 shots which are 25
m separated from each other. The first shot was positioned at position 3000 m in the
model (see figure 7.17); the last shot was positioned at 8975 m. A single trace has a
3000
0
4000
5000
lateral position [m]
6000
7000
8000
9000
500
depth [m]
1000
1500
2000
2500
3000
Figure 7.17
Gray scale representation of a part of the Marmousi model. The potential hydrocarbon
reservoir is positioned around x = 7000, z = 2500. Note the complicated structure
above the reservoir.
Chapter 7: Numerical data examples
3000
0
4000
5000
lateral position [m]
6000
7000
157
8000
9000
500
depth [m]
1000
1500
2000
2500
3000
Figure 7.18
Kirchhoff depth migration of the Marmousi model. Note the poor image quality at the
turtle back (x = 600, z = 2500) due to the complex overburden.
length of 4 s and a sampling interval of 4 ms. The potential hydrocarbon reservoir
is positioned around x = 7000, z = 2500 and can be recognized by the horizontal
levels in the turtle-back structure. This objective was modeled as a structural high
representing a hydrocarbon-bearing sand with a large acoustic impedance contrast.
On top of the reservoir, above the unconformity, a complicated over thrust structure
is present. This over thrust structure will distort the wavefields, which are emitted
at the surface, at the target zone significantly and are the most important reason
why it is difficult to get a good image of the target zone.
The pre-processing for the Marmousi data set was carried out in a two step approach (described in Rietveld, 1995, p. 56) in which the surface related multiples
and the thin layer reverberations were removed. The near offsets were interpolated
using a CMP interpolation technique. This pre-processed data set, including the
interpolated offsets, is used as the input for the different migration algorithms. The
migration results shown in this chapter are carried out with three different methods;
Kirchhoff depth migration, recursive depth migration and CFP imaging.
Kirchhoff migration
The Kirchhoff depth migration result shown in figure 7.18 is obtained from the
internet and is shown here to be able to make a comparison possible with the other
migration methods. The imaged result is obtained with a computational efficient
implementation of a 2-D Kirchhoff pre-stack depth imaging algorithm written by Joe
Matarese (MIT). The imaging result is obtained by using first-arrival traveltimes for
the true velocity model. Note that due to the complex overburden the image quality
in the middle of the image, below z = 2500 m, is poor.
158
7.6 Comparison of imaging results for realistic numerical data
3000
0
4000
5000
lateral position [m]
6000
7000
8000
9000
500
depth [m]
1000
1500
2000
2500
3000
Figure 7.19
Recursive depth migration of the Marmousi model. The turtle back structure is well
imaged, however, due to the presence of shadow zones (e.g. at x = 5500, z = 2500)
some parts of the model cannot be imaged very well.
Recursive depth (F-X) migration
The recursive depth migration is carried out with the WLSQ extrapolation operators
as described in appendix A. For the inverse extrapolation of the surface data a depth
step of 12.5 m, an extrapolation operator length of 25 points, a maximum design
angle of 65◦ and a maximum frequency of 55 Hz was used. The recursive depth
migration result is shown in figure 7.19. In this imaging result the deeper part is
imaged very well and the turtle back structure is clearly visible. The great accuracy
of the F-X migration makes it a useful standard for comparison with the results
obtained from other imaging methods. The F-X migration result clearly shows
the faults in the upper part of the model and the image objective; the flat event
centered at a depth of about 2440 m at x = 6000. The better image quality of
the F-X migration compared with the Kirchhoff depth migration is due to the fact
that recursive depth extrapolation automatically handles energy multi pathing from
upper surface points to depth points, while Kirchhoff migration allows a few paths
at most (in the result shown in figure 7.18 only one) to connect an upper surface
point with a depth point.
CFP imaging
For the CFP imaging the time-domain implementation of the CFP method was used
(see appendix C) and only the first arrival times of the focusing operators were used.
The traveltime operators were calculated by making use of a scheme based on the
work of Vidale (1988) and a smoothed version of the original velocity model. For
the first focusing step the focus points were chosen at every 50 m starting at depth
Chapter 7: Numerical data examples
one-way image time [s]
3000
0
4000
5000
lateral position [m]
6000
7000
159
8000
9000
0.4
0.8
1.2
Figure 7.20
CFP one-way time image of the Marmousi model. Due to the complex overburden the
focusing operator for the deeper part of the model are calculated erroneously which
results in a distorted image of the layers below z = 2500.
level z = 50 m, which means that 60 operators are used for every lateral position.
A coarser operator grid in the first focusing step gives an inferior result due to the
many details in the model. The CFP image in one-way image time is shown in figure
7.20 and the depth image is shown in figure 7.21. Due to the complex overburden
the focusing operator for the deeper part of the model are calculated erroneously in
the used ray-tracing scheme which results in a distorted image of the layers below
z = 2500. There are several reasons why the focusing operators are calculated
erroneously:
• taking only the first arrival times into account can give an operator time which
belongs to a ray which travels along a high velocity layer of the model (head
wave), this ray contains less energy than the ray traveling through the layers.
• the complex overburden gives rise to timing errors in the traveltime calculation
program. Vidale (1988) showed that for certain areas an error of 5-10% is
possible.
Geoltrain and Brac (1993) have also argued that it is difficult to image Marmousi
with only first arrival times obtained with an Eikonal solver. They have used Kirchhoff depth migration in conjunction with traveltimes computed by finite-differencing
the Eikonal equation. Their disappointing results did not originate because of the
intrinsic limitations of Kirchhoff migration, but rather from the failure of finitedifferencing to compute traveltimes representative of the energetic part of the wave
field. The consequence is that energetic seismic events are imaged with an incorrect
operator and turn out mis-positioned. The solution for this problem is using a better
160
7.6 Comparison of imaging results for realistic numerical data
3000
0
4000
5000
lateral position [m]
6000
7000
8000
9000
500
depth [m]
1000
1500
2000
2500
3000
Figure 7.21
CFP depth image result of the Marmousi model. Note that below z = 2500 the layers
in the middle are visible, but imaged at the wrong position due to errors in the used
focusing operators.
ray-tracer which calculates traveltimes that correlate very well with the energetic
events in the correct focusing operators (Gray and May, 1994).
To investigate the image quality of the result obtained with the CFP method the
coherency measurement, as introduced in the chapter 5 page 79, is used; the result of
this measurement is shown in figure 7.22. This normalized coherency measurement
gives an indication how the different events in the CFP image gathers are aligned
before summation to the final image trace. If correct operators are used all events
should align and give a high (black in figure 7.22) coherency measurement. The
coherency measurement shows clearly where erroneous focusing operators have been
used; the middle lower part of the model.
The coherency measurement gives an indication of the alignment in a CFP image
gather. In figure 7.23 two image gathers are shown; one for a position at x = 4000
and one for a position at x = 6000 in the model . The gather taken at x = 4000
represent a part of the model which is not very complicated, this is observed in the
good alignment of the events in the image gather, meaning that the used traveltime
operators described the energetic part of the wavefront correctly. The other image
gather at x = 6000, shows that the alignment of the deeper events is poor, meaning
that the first arrival times of the used traveltime operators were not able to describe
the energetic part of the wavefront correctly. Note that observed differences in
the alignment of the traces in the CFP image gather are well represented by the
coherency measurement of figure 7.22
Chapter 7: Numerical data examples
one-way image time [s]
3000
0
4000
5000
lateral position [m]
6000
7000
161
8000
9000
0.4
0.8
1.2
Figure 7.22
Coherency measurement of the CFP image gathers used in the construction of the oneway time image. In the figure the color black (coherency value of 0.7) means a good
coherency and white (0.0) means no coherency. The measurement should be compared
with the one-way time image of figure 7.20.
lateral position [m]
4000
5000
0.4
0.8
1.2
lateral position [m]
6000
7000
0.4
0.8
1.2
a. image gather for x = 4000
Figure 7.23
5000
0
one-way image time [s]
one-way image time [s]
3000
0
b. image gather for x = 6000
Image gathers for two different lateral positions in the model. The image gather for
the less complicated part of the model (a), shows a good alignment of the events. The
image gather taken in the more complicated part of the model (b) is not as well aligned.
162
7.6 Comparison of imaging results for realistic numerical data
3000
0
4000
5000
lateral position [m]
6000
7000
8000
9000
time [s]
0.5
1.0
1.5
2.0
0
500
depth [m]
1000
1500
2000
2500
3000
0
500
depth [m]
1000
1500
2000
2500
3000
Figure 7.24
Focusing beam and snapshot for a focusing operator at x = 6000, z = 2500 in the
Marmousi model. The used traveltime operators are superimposed as a white line on
the top picture.
Chapter 7: Numerical data examples
one-way image time [s]
one-way image time [s]
3000
4000
5000
lateral position [m]
6000
7000
163
8000
9000
0.8
1.0
1.2
0.8
1.0
1.2
Figure 7.25
Coherency measurement (bottom) of the CFP imaging result (top) based on operators
calculated in a 1-dimensional macro model. Note the low coherency compared with
figure 7.22, both coherency measurements are displayed with the same scale factor.
In figure 7.24 a focusing operator together with its snapshot and focusing beam is
shown to illustrate how complicated the propagation through the overburden can get.
The wavefront at the surface, observed in the top picture, is completely broken due
to the propagation through the complex overburden in the model. The beam shows
clearly the illumination area for the defined focus point, during the propagation
several other areas are focused and de-focused due to the complex structure in the
model. To generate the results in figure 7.24 use was made of the extrapolation
operators as defined in appendix A. To be able to use time-domain methods the
most energetic part of the wavefront shown in figure 7.24 must be approximated in
an efficient way by a traveltime operator. The used Eikonal solver is not designed
for this task, which is observed in the white line in the top picture representing the
calculated traveltimes used in the imaging.
By using a 1-dimensional model for the generation of the operators the sensitivity
of the coherency measurement to a wrong macro model can be tested. In the 1dimensional velocity model a linear velocity gradient is defined with c = 1500 m/s
at z = 0 m and c = 3660 m/s at z = 3000 m. The focus points are only defined
for the deeper part of the model from z = 2000 m to z = 3000 m with a depth
step of 50 m. The imaged result is shown in figure 7.25. Note that the events are
positioned at the same one-way image time as in the result obtained with the correct
model (see section 6.4 on page 132), a map in depth will give an erroneous position.
The left part of the model is less complicated and gives a reasonable image quality
and coherency measurement, however, the middle and right part of the coherency
measurement show that the image quality is poor.
164
7.6 Comparison of imaging results for realistic numerical data
0
0.1
0.2
0.3
0.4
0.5
0.6
lateral position [m]
0.7
0.8
0.9
x10 4
1.0
1.1
1.2
1.3
1.4
1.5
0
500
depth [m]
1000
1500
2000
2500
3000
3500
Figure 7.26
7.6.2
Profile A-A’ from the SEG-EAGE salt model crosses many of
the more difficult structural elements in the model.
For a
complete overview of the 3D model the reader is referred to
http://sepwww.stanford.edu/seg/research/3Dmodel/SALTHOME/segsalt.html.
SEG/EAGE salt dome model
The recently developed SEG-EAGE salt model was built to address data quality
issues encountered around the types of salt features found on the US Gulf Coast
(Aminzadeh et al., 1995). A part of the modeled data, generated for a 2D slice
along profile A-A’ of this model (see figure 7.26), has been used by O’Brien and Gray
(1996) to compare migration results of Kirchhoff and F-X migration. At this point
I would like to thank Mike O’Brien and Walter Rietveld, of Amoco Exploration and
Production Technology Group in Tulsa, for providing me with the modeling results
along the 2D profile. The modeled data consists of 325 shot records moving from
left to right through the model. The first shot has its shots position at x = 0, and
the receiver line belonging to this first shot starts at x = −4267.2 m (= −14000 ft)
and ends at x = 0. The next shot, and its receiver line belonging to it, is positioned
48.768 m (=160 ft) at the right of the first shot. The receiver sampling interval is
24.384 m (= 80 ft), the time sampling is 8 ms and the maximum frequency present
in the data is about 25 Hz. The reflectors in the model are modeled by spikes
(with a velocity 120 % of the background velocity) positioned in a smoothly varying
background, the salt structure is presented by a high velocity (c = 4480 m/s) area
as shown in figure 7.26.
Figure 7.27 shows the zero-offset section extracted from the shot records. The reflections and diffractions related to the salt body are very complicated and are a
real challenge for the migration algorithms. In the zero-offset section the base of the
salt, and the structures below it, are not clearly visible. A lot of diffraction curves,
Chapter 7: Numerical data examples
0
0
0.1
0.2
0.3
0.4
0.5
0.6
lateral position [m]
0.7
0.8
0.9
165
x10 4
1.0
1.1
1.2
1.3
1.4
1.5
two-way time [s]
1
2
3
4
5
Figure 7.27
Zero-offset section extracted from the modeled shot records. Note the complicated area
right from the salt dome at x = 12000 and t = 2.0. An exponential taper is used to
display the deeper events better.
originating from the high velocity contrast of the salt body, are observed instead.
The CFP imaging result in one-way time is shown in the top picture of figure 7.28
the CFP depth result is shown in the bottom picture of figure 7.29. The CFP images
are obtained by placing a focus point at every 100 m in depth, starting at 25 m, for
all receiver positions, meaning that in total 38 operators are used for every lateral
position. The quality of this image is not as good as F-X migration result due to the
problems with the used traveltime modeling program, which provided the traveltimes
for the focusing operators. For a comparison with a Kirchhoff depth migration result
the reader is referred to O’Brien and Gray (1996). The coherency plot belonging to
the one-way image is shown in the bottom picture of figure 7.28, where it is observed
that below the salt body the coherency value is very low, indicating that the used
Eikonal operators are not accurate enough.
The top picture in figure 7.29 shows the F-X migrated section. This image is considered as a reference image for the other migration results. The depth step used in
the extrapolation algorithm was 12 m. For the extrapolation carried out in the migration the WLSQ operators (see appendix A) were used with a convolution length
of 25 points and a maximum propagation angle of 65◦ . To be able to image energy
for the large angles the shot records are positioned into the velocity model with some
additional zero-traces (150 traces) added on both sides of the shot record. In this
way the data in the shot record is allowed to migrate outside its own acquisition
aperture. Below the salt body several structures, like the steep fault planes, are not
reconstructed because the energy of these reflectors did not reach the receiver at
the surface (due to a limited acquisition aperture or the energy from the steep fault
166
7.6 Comparison of imaging results for realistic numerical data
one-way image time [s]
0
x10 4
0
0.1
0.2
0.3
0.4
0.5
0.6
lateral position [m]
0.7
0.8
0.9
1.0
1.1
1.2
1.3
1.4
1.5
0
0.1
0.2
0.3
0.4
0.5
0.6
lateral position [m]
0.7
0.8
0.9
1.0
1.1
1.2
1.3
1.4
1.5
0.5
1.0
1.5
one-way image time [s]
0
x10 4
0.5
1.0
1.5
Figure 7.28
CFP migration result in one-way time (top), the coherency measurement (bottom). In
the migration results all the reflectors are visible. The coherency plot shows that the
coherency is large at the boundaries in the model, indicating a good alignment of the
image gathers. Note that below the salt dome the illumination is poor.
planes was trapped in the model and did not reach the surface at all). Note that
the computation time of the CFP image (including the time of the calculation of the
operators) is about 10 times faster than the F-X migration.
In figure 7.30 a focusing operator is shown together with its snapshot and focusing
beam. The wavefront measured at the surface is not as broken as the operator for the
Marmousi model shown in figure 7.24. In the operator picture (top) it is observed
that the first arrival times for this operator position (white line) are in most cases
Chapter 7: Numerical data examples
0
167
x10 4
0
0.1
0.2
0.3
0.4
0.5
0.6
lateral position [m]
0.7
0.8
0.9
1.0
1.1
1.2
1.3
1.4
1.5
0
0.1
0.2
0.3
0.4
0.5
0.6
lateral position [m]
0.7
0.8
0.9
1.0
1.1
1.2
1.3
1.4
1.5
500
depth [m]
1000
1500
2000
2500
3000
3500
0
x10 4
500
depth [m]
1000
1500
2000
2500
3000
3500
Figure 7.29
F-X migration (top) and the CFP depth migration (bottom) result of the SEG salt dome
model. The flat layers below the salt structure are very well imaged. At the bottom of
the image not all the steep faults present in the model are imaged.
time coincident with the wavefront which contains most of the energy. Note that
only for the higher offsets on the left-hand side the calculated traveltime deviates
from the first arrival times defined by the wavefront. This means that by taking into
account only the first arrivals in the CFP imaging algorithm, it should possible to
obtain a good image of the subsurface. The snapshot in figure 7.30 (middle picture)
shows how the wavelet is stretched due to the high velocity in the salt dome.
168
7.6 Comparison of imaging results for realistic numerical data
0
0
0.1
0.2
0.3
0.4
0.5
0.6
lateral position [m]
0.7 0.8 0.9
x10 4
1.0
1.1
1.2
1.3
1.4
1.5
1
time [s]
2
3
4
5
0
500
depth [m]
1000
1500
2000
2500
3000
3500
0
500
depth [m]
1000
1500
2000
2500
3000
3500
Figure 7.30
Focusing beam and snapshot for a focusing operator at x = 9000, z = 3500, below
the salt body in the SEG salt model. The wavefront which contains most of the energy
coincide for a great part with the first arrivals used in the CFP imaging method.
Chapter 8
Field data examples
In this chapter two field data sets, which are shown with the permission of the
sponsors of the DELPHI consortium who provided the data, are used to show how
the CFP method performs on field data. All field data examples presented in this
chapter represent 1-dimensional marine acquisition lines. The first data set discussed
is from the Northern part of the North sea and does not contain many complicated
structures. The second data set includes the reflection of a salt body, which makes
it an interesting data set from an imaging point of view.
8.1
Mobil
The field data set used in this section has been provided by Mobil Research and
Development Corporation and has been made available for the purpose of testing
seismic inversion methods and comparing them in a workshop. The data set contains
marine seismic data and two well log measurements within the seismic line. The data
has been acquired above the North Viking Graben in the North Sea and contains
an area with hydrocarbon reservoirs. The primary reservoir objectives in the North
Viking Graben are Jurassic clastic sediments. The reservoirs in these sediments are
sometimes vertically stacked and separated by shales. Hydrocarbon traps are usually
fault-bounded structures, but some traps are associated with the unconformity at
the base Cretaceous. This unconformity will become visible in the one-way time
image at approximately 1.0 seconds.
The seismic line consists of 1001 shot records, oriented in a structural dip direction,
each shot record contains 120 channels. The shot point interval and receiver group
interval are 25 m. The seismic data were acquired in approximately 350 meters of
water, meaning that there are strong surface related multiples present in the data.
To remove these surface related multiple the method of Verschuur et al. (1992) was
used.
For the imaging, the shot numbers 354 to 1053 have been selected, with 1024 samples
170
8.1 Mobil
per trace. The pre-processing steps carried out on these shots consists of
• direct wave mute.
The direct wave is muted out of the data because is does not contain any
information of the subsurface of the earth.
• 3D to 2D spherical correction.
A simple gain has been applied to simulate line source instead of point source
responses. After this correction 2-dimensional processing schemes can be used.
• replacing bad traces.
The bad traces were removed from the shot record and re-interpolated from
the good traces using rough NMO correction and spline interpolation. The
tradeoff being made here is that the influence of a bad trace on the image
quality is considered to be much worse than the influence of an interpolated
trace.
• wavelet deconvolution
A predictive deconvolution has been applied with a gap of 20 ms and a filter
length of 240 ms. This deconvolution is applied to get a better multiple removal
in the next processing step.
• receiver sensitivity correction.
The receivers showed a consistent sensitivity behavior through the different
shot records. Least-squares inversion techniques were used to correct these
amplitude fluctuations. The fluctuations present in the sources have been left
in the data. Unbalanced traces give rise to a lower quality of the image.
• interpolation of missing shots and near offsets.
Application of the surface-related multiple elimination method requires a full
coverage of the shot and receivers up to zero-offset. In the selected shot range
6 shot records were missing, and have been inserted and interpolated in the
common offset plane. Approximately 10 near offset traces were missing in
each shot. They have been created by using the parabolic Radon transform to
extrapolate the data to zero-offset in the CMP domain (Kabir and Verschuur,
1993). These interpolated traces are not used in the imaging.
After these basic pre-processing steps the multiple elimination step has been carried
out. The surface related multiple elimination method is based on wave theory.
Verschuur et al. (1992) have proven that by taking temporal and lateral autoconvolutions of seismic data, an accurate prediction of the surface related multiples
can be obtained. Subsurface information is not needed, but information on the free
surface reflectivity (assumed to be -1 for marine data) and the source wavefield is
required. As the latter is not known in practice, the method is applied adaptively:
by eliminating the multiples the source signature is estimated as well.
An initial macro model was estimated with the areal shot record technology as
Chapter 8: Field data examples
2.4
0
B
2.0
1.8
lateral position [m]
1.6
1.4
1.2
A
x10 4
1.0
0.8
1475
1475
2050
1870
2020
1000
2235
2250
1500
2340
2300
2000
2600
3000
3250
500
depth [m]
2.2
171
2500
3100
3000
3375
3500
3650
3500
4000
Figure 8.1
The estimated macro model for the Mobil data set. The line indicate the positions of the
wells, indicated with A and B, where the CFP analysis is done.
proposed in Rietveld and Thorbecke (1994). Last year (1996) this macro model was
updated by using the CFP move-out analysis (Kabir and Verschuur, 1996). The
updated estimated macro model consists of seven macro layers as shown in figure
8.1. By using the same technology for the estimation of the macro model and the
imaging, the macro model estimation and the imaging are closely connected to each
other and model errors observed in the construction of the image can be directly
used to update the macro model.
The data set contains two well positions, indicated in figure 8.1, within the seismic
line. The CFP analysis presented in this chapter is concentrated on these two well
positions. For a first analysis focusing beams are calculated for focus points at every
macro boundary below the two well positions A (at position x = 11700 m) and
B (at position x = 21250 m). In figure 8.2 the focusing beams and the focusing
operators are shown. The beams show the focusing of energy, for the different
synthesis operators, at the defined macro boundaries. Note that in between the
boundaries the illumination energy will not be sufficient to image the events in
between the boundaries correctly. So for a good second focusing step more operators
are needed as used in figure 8.2. The combination of beams also show that close to
the surface the energy in the beam is reduced due to the used extrapolation operator
and the energy for the deeper part of the model is reduced due to the limited aperture
range of the acquisition. For a better analysis of the illumination energy different
beams were calculated where the distance between the focus points was changed.
From these experiments it is concluded that for ∆z = 150 m a sufficient illumination
of the subsurface is obtained.
An X-gather is defined at the position of well A and is constructed with synthesis
172
0
8.1 Mobil
x10 4
lateral position [m]
1.1
1.3
0.9
0
1.0
1.5
4000
a. X-operator at well A
1.9
lateral position [m]
2.1
b. X-beam at well A
x10 4
2.3
0
0.5
2.0
lateral position [m]
2.2
x10 4
1000
depth [m]
time [s]
2000
3000
2.0
0
x10 4
1000
depth [m]
time [s]
0.5
lateral position [m]
1.1
1.3
0.9
1.0
1.5
2000
3000
2.0
c. X-operator at well B
Figure 8.2
4000
d. X-beam at well B
Focusing operators and focusing beams for focus points defined at every macro boundary
at the lateral position of the wells. The information in the beams is used to determine the
focus point density in the first focusing step.
Chapter 8: Field data examples
0
0
500
offset [m]
1000
1500
0
500
offset [m]
1000
1500
0.5
one-way time [s]
one-way time [s]
0.5
0
173
1.0
1.5
1.0
1.5
2.0
2.0
a. X-gather for well A
Figure 8.3
b. CMP gather for well A
X-gather and CMP gather at the position of well A. Note that the first 10 traces in the
CMP gather are interpolated traces.
operators defined for focus points with a sampling rate of ∆z = 150 m (∆T ≈ 150
ms). Figure 8.3 shows the X-gather and the CMP gather defined for the same
lateral position. The CMP gather is displayed with the same time sampling and
offset range as the X-gather. Note that the first 10 traces (until offset 250 m) in
the CMP gather are interpolated traces, the X-gathers are constructed without the
interpolated traces. In figure 8.3 it is observed that for the higher offsets the X-gather
shows more continuous events than the CMP gather. The two main differences
between the CMP gather and the X-gather are: ➊ the CMP gather has its midpoint defined at the surface, while for the X-gather the focus points are defined at
different depth levels in the subsurface at the same lateral position. ➋ the traces in
the X-gather are constructed by means of a synthesis step and are the result of a
Fresnel summation, the traces in the CMP gather are selections from the traces of
the shot records.
The constructed X-gathers are the input for the second focusing step. In figure 8.4 for
both well positions A and B the move out corrected X-gathers (CFP image gathers)
are shown with focusing operators defined at every 50 ms of the one-way image time.
The interpolated traces are not used in the calculation and an overlapping area of 10
sample points is used between two adjoined CFP gathers. In the CFP image gathers
the good alignment of the events is observed. Note that for the shallow part more
174
0
8.1 Mobil
0
500
offset [m]
1000
1500
0
500
offset [m]
1000
1500
0.5
one-way time [s]
one-way time [s]
0.5
0
1.0
1.5
1.0
1.5
2.0
2.0
a. image gather well A
Figure 8.4
b. image gather well B
CFP image gather at the positions of well A and well B where at every ∆T = 50 ms a
focusing operators is defined for the first focusing step.
operators are needed than for the deeper part. It is therefore more efficient to use
operators which are differently sampled in one-way image time; dense sampled close
to the surface and more sparsely sampled for the deeper parts. However if one is
only interested in the deeper parts of the image the focusing operators at the surface
need not to be calculated at all.
The CFP migration result is shown is figure 8.5. The migration result is obtained
with synthesis operator sampled with approximately 80 ms in single vertical time,
which is equivalent with a depth step of ∆z = 150 m. Due to the tube like shape of
the focusing beams this focus point sampling in the first focusing step is sufficient to
obtain a good image of the subsurface. To build the one-way time image a total of
19896 operators were used in the first focusing step (24 per lateral position for 829
lateral positions). The interpolated near offset traces are not used in the imaging
step, because these artificial traces give a strong disturbing contribution to the final
imaging result.
The image shown in figure 8.5 shows an unconformity at 1.1 seconds. Above the
unconformity the layers are well imaged and there are some overlapping structures
visible below x = 22000 m at 0.6 seconds. Below the unconformity several block
structures, for example at x = 13000, t = 1.2, are present. These block structures
could contain hydrocarbon reservoirs.
0
2.4
2.3
2.2
2.1
2.0
1.9
1.8
1.7
lateral position [m]
1.6
1.5
1.4
1.3
1.2
1.1
1.0
0.9
0.8
x10 4
0.7
1.0
Chapter 8: Field data examples
one-way image time [s]
0.5
1.5
175
Figure 8.5 Prestack migration result obtained with CFP migration method. Note that only a limited number of operators is used in the first focusing
step to construct the complete image: ∆T = 80 ms. An exponential gain in the time-direction was used to amplify the weaker events in
the image.
0
2.4
2.3
2.2
2.1
2.0
1.9
1.8
1.7
lateral position [m]
1.6
1.5
1.4
1.3
1.2
1.1
1.0
0.9
0.8
x10 4
0.7
depth [m]
8.1 Mobil
1000
2000
3000
176
4000
Figure 8.6 Prestack depth migration result obtained with recursive F-X migration. The depth step in the recursive extrapolation was chosen 10 m
and the WLSQ extrapolation operators, as discussed in appendix A, were used.
Chapter 8: Field data examples
one-way image time [s]
0
2.4
2.2
2.0
1.8
lateral position [m]
1.6
1.4
177
x10 4
1.2
1.0
0.8
0.5
1.0
1.5
Figure 8.7
Coherency measurement of the one-way time image. Black indicates a high coherency,
meaning that the event in the CFP image gather is well aligned.
The F-X migration result is shown in figure 8.6, here the same kind of structures, as
in the one-way time image, are observed . However there are also some differences;
just above the unconformity, from x = 16000 to x = 23000, the one-way time
image seems to have resolved the structures better. Below the unconformity the
F-X migration shows the blocks better, but with field data it remains difficult to tell
which image is the best.
Figure 8.7 shows the coherency measurement of the image shown in figure 8.5. In
the coherency measurement it is observed that the layers close to the surface can be
imaged better by making use of a denser sampling in depth for the focus points in
the first focusing step. Below 1.5 seconds the alignment in the image gathers is poor,
indicating wrong focusing operators, a low signal to noise ratio or the absence of a
reflector. Figure 8.8 shows the steps to be taken to extract reflectivity information
for one focus point. First a CFP gather is calculated for the focus position of interest.
The quality of this CFP gather is checked by means of a move-out panel, if the event
aligns at t = 0 then the used operator gives a good description of the propagation
properties. If the panel does not indicate a good alignment the operator needs to
be updated until a good alignment in the move-out panel is the result. In the panel
shown in figure 8.8b it is observed that there is a reflection event below t = 0 s
which is flat, indicating that the used focusing operator is not perfect. Next a 2dimensional cross-correlation between the CFP gather and the focusing operator can
be carried out to obtain the reflectivity gather as shown in figure 8.8c. Around t = 0
and x = 0 this reflectivity gather contains the reflection information of the defined
focus point measured at the surface. Using a τ − p transformation on this part of
the reflectivity gather gives the reflectivity function of the focus point in terms of
the ray-parameter shown in figure 8.8d.
178
8.1 Mobil
2.0
0
lateral position [m]
2.2
x10 4
2.0
-1.0
0.5
-0.5
delay [s]
time [s]
x10 4
lateral position [m]
2.2
1.0
1.5
0
0.5
2.0
1.0
a. CFP gather for x = 21250, z = 3075
offset [m]
2.0
b. move-out panel of (a)
x10 4
2.2
-0.3
-0.6
-1.0
-0.2
ray-parameter [ms/m]
-0.1
0
0.1
0.2
0.3
-0.4
-0.5
intercept time [s]
delay [s]
-0.2
0
0
0.2
0.5
0.4
0.6
1.0
c. reflectivity gather
Figure 8.8
d. reflectivity function
The gathers involved in the calculation of a reflectivity function for one point in the
subsurface. The analysis is carried out for a focus point at the position of well B at a
depth z = 3075 m.
Chapter 8: Field data examples
1.6
0
1000
1.8
2.0
lateral position [m]
2.2
2.4
depth [m]
3000
4000
x10 4
2.6
2.8
3.0
1835
1790
2260
2190
2233
2415
2000
179
4120
4270
4150
4480
4200
3200
4000
3945
3850
4080
4200
4200
5000
Figure 8.9
8.2
The estimated macro model for the ELF data set. The model consists of 7 layers which
are estimated by making use of CFP move-out panels.
ELF
The data used in this section was originally owned by ELF and has been made available by Institut Français du Pétrole (IFP) for the use in the Integrated Structural
Imaging project (ISI project) of the Joule II program of the European Committee.
Below the acquisition surface a complicated salt structure is present which is probably triggered by faulting of the basement. The left flank of the salt structure can
be an overhang and is an interesting area for imaging. The marine line used in this
thesis is one out of two streamer lines of the same shot acquisition. The used line
contains 378 shot records with an offset range from -187.5 m (trace 1) to -3362 m
(trace 120). The receiver spacing is 26.66 m, the shot spacing 40 m and every trace
contains 1250 samples sampled with 4 ms. The first shot position starts in the model
at position 16020 m and the last shot position end at 31100 m. The macro model
shown in figure 8.9 is used in the recursive depth migration and for the modeling of
the focusing operators used in the CFP migration. The model consists of 7 layers
which are estimated by making use of CFP move-out panels (Kabir, 1997).
As a first experiment on this data set CFP gathers are constructed for positions left
and right from the estimated position of the salt structure. In figure 8.10 move-out
corrected CFP gathers are shown with a focus point defined at a depth of 2500 m
in the macro model (approximately 1 second one-way image time) for x = 22000 m.
The difference between the two CFP gathers shown in figure 8.10 is that the left
CFP gather is constructed for focusing in detection and the right one for focusing
in emission, but both gathers are defined for the same focus point. For focusing in
detection the move-out corrected CFP gather is constructed at the source positions
180
8.2 ELF
-2000
0
-1000
CFP offset [m]
0
1000
2000
-2000
0
CFP offset [m]
0
1000
2000
0.5
time [s]
time [s]
0.5
-1000
1.0
1.5
1.0
1.5
2.0
2.0
a. CFP gather for detection
Figure 8.10
b. CFP gather for emission
CFP gathers for a position left of the flank of the salt dome (x = 22000 m). Figure a
show the CFP gather for focusing in detection and b shows the CFP gather for focusing
in emission.
with a sampling interval of 40 m. Note that due to the used acquisition geometry
flat events should occur at positive CFP offsets. Any events occuring at the negative
offsets indicate the occurrence of non flat reflectors. The CFP gather for focusing in
emission, sampled at the half of the receiver sampling, 13.33 m, shows the alignment
of the events at the negative CFP offsets. At this side of the flank of the salt structure
it is not expected that with the used acquisition geometry much information is
measured from the flank of the salt structure. Note that the difference between the
two gathers is that for focusing in emission the remaining traveltime is defined from
the focus point to the receiver position at the surface, while for focusing in detection
the remaining traveltime is from the focus point to the source position. Note also
that for the construction of one trace out of the CFP gather for focusing in detection
120 traces (receivers per shot position) are used while for focusing in emission only
40 traces (sources per receiver position) are used.
Choosing a focus point on the other side of the flank of the salt structure, at x =
26000 m and z = 2500 m, it is observed in figure 8.11 that the CFP gathers contain
more information about the right side of the salt structure. This information is
clearly visible in the CFP gather for focusing in emission at the positive offsets. By
choosing more focus points on this side of the flank and combining the information in
the calculated CFP gathers a good image can be obtained from the flank of the salt
Chapter 8: Field data examples
-2000
0
-1000
CFP offset [m]
0
1000
2000
-2000
0
CFP offset [m]
0
1000
2000
0.5
time [s]
time [s]
0.5
-1000
181
1.0
1.5
1.0
1.5
2.0
2.0
a. CFP gather for detection
Figure 8.11
b. CFP gather for emission
CFP gathers for a position right of the flank of the salt dome (x = 26000 m). Figure a
show the CFP gather for focusing in detection and b shows the CFP gather for focusing
in emission.
structure as shown by Alá’i (1997). Note that due to the sampling of the sources at
40 m, the construction for focusing in emission is done with a relative large sampling
interval, giving rise to aliased results.
The image result of the recursive pre-stack depth migration, with a depth step of 10
m, is shown in figure 8.12. In this image the right flank of the salt structure is visible, the left flank is not as clear due to the, already discussed influence, acquisition
geometry. A deep synclinal structure is observed around the position x = 27000 m
and z = 3000 m. The overhang which could be present in the structure for geological reasons is not observed in the obtained depth image. Note that the possible
information of the overhang cannot be imaged with the used pre-stack depth migration method scheme because of the use of one-way propagation operators, which
are not able to take multiple reflections from one layer into account (Wapenaar and
Berkhout, 1989).
Figure 8.13 shows the CFP migration result where a reflection, which could originate
from the bottom of the salt body, can be observed at x = 24000 and t = 1.2 s. Note
that this reflection is not as clear present in the recursive depth migration result.
At the bottom of the image, after 1.5 seconds, some linear artifacts are visible
related to the interpolation of the focusing operators in the second focusing step.
The coherency measurement of figure 8.14 indicates good alignment of the events
1.6
0
1.7
1.8
1.9
2.0
2.1
2.2
lateral position [m]
2.3
2.4
2.5
2.6
2.7
2.8
2.9
3.0
x10 4
3.1
1000
depth [m]
8.2 ELF
2000
3000
4000
5000
6000
182
Figure 8.12
Prestack depth migration result obtained with F-X migration. The depth step in the recursive extrapolation was chosen 10 m and the
WLSQ extrapolation operators as discussed in appendix A were used.
1.6
0
1.7
1.8
1.9
2.0
2.1
2.2
lateral position [m]
2.3
2.4
2.5
2.6
2.7
2.8
2.9
3.0
x10 4
3.1
Chapter 8: Field data examples
one-way image time [s]
0.5
1.0
1.5
2.0
Figure 8.13
183
Prestack migration result obtained with the CFP migration method. Note that only a limited number of operators is used in the first
focusing step to construct the complete image: ∆T = 80 ms.
184
8.2 ELF
1.6
0
1.8
2.0
lateral position [m]
2.2
2.4
x10 4
2.6
2.8
3.0
depth [m]
0.5
1.0
1.5
2.0
Figure 8.14
Coherency measurement of the one-way time image. Black indicates a high coherency,
meaning that the event in the CFP image gather is well aligned.
between 1 and 1.5 seconds. Inside the salt structure the image quality is poor except
for the event starting at x = 24000 m and t = 1.2 s. The top part of the image
shows a low coherency value caused by the chosen focus point distribution.
Finally a CFP stack is shown for focus points defined at depth level z = 3500 m and
a lateral distance between the focus points of 26.66 m. The CFP stack gives a good
indication of the events around the defined depth level.
1.6
0
1.8
2.0
lateral position [m]
2.2
2.4
x10 4
2.6
2.8
3.0
one-way time [s]
0.5
1.0
1.5
2.0
Figure 8.15
CFP stack for focus point defined at depth level z = 3500 m. Note that at approximately
1.3 s the result is well focused.
Appendix A
Operator optimization
In homogeneous media the one-way extrapolation operator in the kx , ky − ω
(wavenumber-frequency) domain is a simple analytical function, called the phase
shift operator and given by (Gazdag, 1978):
r
ω2
− (kx2 + ky2 )∆z)
(A.1)
Ỹ (kx , ky , ω, ∆z) = exp (−j
c2
with ∆z being a small extrapolation step, c the propagation velocity of the medium
and ω the angular frequency. The advantage of using the phase shift operator in the
kx , ky −ω domain is that the desired extrapolation result is obtained by multiplication
of the data with the phase shift operator. But simple multiplication in the kx , ky − ω
domain rules out the possibility of applying a laterally varying operator. Another
disadvantage is the numerical artifact in x due to under sampling in wavenumber
domain. To allow laterally varying medium functions and less numerical artifacts a
convolution operator in the x, y −ω (space-frequency) domain should be used. When
the spatial extrapolation operator is used in an explicit recursive depth migration
algorithm it must be calculated in such a way that it gives reliable and stable results
within a reasonable computation time. To arrive at this goal two steps must be
taken; the first step is an optimum design of the spatial operator and the second
step deals with a fast implementation of the spatial convolution. It turns out that
the most efficient algorithms combine these two steps and design a spatial operator in
such a way that it can be implemented in a fast way. In this appendix extrapolation
operators are derived for 2-dimensional and 3-dimensional media.
A.1
1-Dimensional operators for 2-dimensional extrapolation
There are several ways to obtain a 1-dimensional spatial convolution operator. For
homogeneous media one usually starts with the exact analytical expression in the
wavenumber-frequency domain and transforms this operator back to the spatial domain. In recent years many methods have been developed to do this transformation
186
A.1 1-Dimensional operators for 2-dimensional extrapolation
in an efficient and optimum way. For the one-way wavefield extrapolation operator
Holberg (1988), Blacquière et al. (1989), Hale (1991b) and Nautiyal et al. (1993)
have proposed methods to arrive at spatial operators which are unconditionally
stable in a recursive extrapolation scheme. In subsection A.1.2 of this appendix an
alternative method is presented for an efficient and controlled transformation from
the wavenumber domain back to the spatial domain. This method can be used to
calculate extrapolation operators which are stable and accurate in a recursive depth
migration algorithm. Before this new method is discussed first some examples of extrapolation operators for homogeneous media are given to illustrate the problems in
operator optimization. At the end of this section the proposed method is compared
with other numerical optimization methods by calculating the impulse response of
a recursive depth migration algorithm.
A.1.1
Analytical space-frequency operators
In the kx − ω domain the extrapolation operator is, for a 2-dimensional medium,
given by the familiar phase-shift operator of equation (A.1) with ky = 0
p
(A.2)
Ỹ (kx , ω, ∆z) = exp (−j k 2 − kx2 ∆z),
Ỹ (kx , ω, ∆z) = exp (−jkz ∆z),
with
(p
k 2 − kx2
p
kz =
−j kx2 − k 2
kx2 ≤ k 2
kx2 > k 2
(A.3)
.
(A.4)
in which k is defined as ωc . Note that for wavenumber values larger than k, the wavefield becomes evanescent (exponentially decaying). The analytical inverse Fourier
transform of equation (A.2) is a scaled Hankel function (see Berkhout, 1984):
Y (x, ω, ∆z) =
−jk∆z
(2)
, H1 (kr)
2r
(A.5)
p
(2)
with r = (x2 + ∆z 2 ) and H1 (kr) (= J1 (kr) − jY1 (kr)) the zero order Hankel
function of the second kind.
To illustrate the behavior of the phase shift operator the same parameters as Hale
(1991b) and Nautiyal et al. (1993) are used; ∆z = ∆x = 12.5 m, ω = 40π [radians/s],
512 wavenumber samples and c = 1000 m/s. In figure A.1 the amplitude of equation
(A.5) is given as function of x, with the chosen parameters. In the same figure the
amplitude of the wavenumber spectrum of a truncated Hankel function (dashed line)
is displayed together with the phase shift operator Ỹ (kx , ω, ∆z) (solid line). The
horizontal axis in figure A.1 represents normalized wavenumber cycles ( nπ
N , radians
per sample in the x direction; e.g. 0.25 is equivalent to the wavenumber kx = k and
represents propagation at 90◦ and 0.1 is equivalent to an angle of 36◦ ). The number
amplitude
Appendix A: Operator optimization
0.5
1.0
0.4
0.8
0.3
0.6
0.2
0.4
0.1
0.2
0
187
0
0
250
500
operator length [m]
750
0
0.1
0.2
0.3
0.4
0.5
cycles
Figure A.1 Analytical spatial extrapolation operator (left) and its wavenumber spectrum (right).
The dashed line is related to the truncated analytical operator (121 points) and the
solid line to the phase shift operator. Note the excellent agreement; for 121 points the
truncation error is negligible.
amplitude
of samples points in the spatial domain is chosen at 121. In all figures shown in this
section only the positive values for kx and x are displayed because of the symmetry
of the operators, meaning that in figure A.1 only 61 x−positions are shown. From
figure A.1 it can be seen that the Hankel function is a long operator and is therefore not very well suited for recursive depth extrapolation. From a computational
point of view long spatial operators are not desired because multiplication in the
wavenumber-frequency domain is replaced by a convolution in the space-frequency
domain. Further, the locally homogeneous assumption in inhomogeneous media is
for a long operator pushed to its limits. If we truncate the Hankel function to a
more suitable number of points (for example 39 points) then it is observed in figure A.2 that the wavenumber spectrum is not stable for a wavenumber close to k.
Recursive application of this operator in a homogeneous medium causes the waves,
traveling with the wavenumber related with the deviation above the amplitude = 1
line, to be amplified at every extrapolation step which can in the end ’blow up’ the
extrapolation result.
0.5
1.0
0.4
0.8
0.3
0.6
0.2
0.4
0.1
0.2
0
➘
0
0
50
100
150
operator length [m]
200
250
0
0.1
0.2
0.3
0.4
0.5
cycles
Figure A.2 Same illustration as in figure A.1, but now with the number of points reduced to 39
points. Note the unacceptable truncation effect, indicated by the arrow in the right
picture.
amplitude
188
A.1 1-Dimensional operators for 2-dimensional extrapolation
0.5
1.0
0.4
0.8
0.3
0.6
0.2
0.4
0.1
0.2
0
0
0
50
100
150
operator length [m]
200
250
0
0.1
0.2
0.3
0.4
0.5
cycles
Figure A.3 Same illustration as in figure A.2, but now with a spatial Gaussian taper applied on
the spatial operator. Note the strong amplitude decay at the higher propagation angles
(> 0.2).
A solution for the unstable behavior of the truncated operator is given by Nautiyal
et al. (1993). They argued that accuracy is important for any numerical method
but stability is crucial because an unstable, accurate method is even less useful
than a stable but inaccurate one. They propose to taper the spatial wavelet with
a Gaussian taper (which guarantees a stable extrapolation operator). To illustrate
the method a Gaussian taper is applied to the operator shown in figure A.2. The
πN
chosen Gaussian taper decays from a value of 1 at x = 0 to a value of cos2 ( 2(N
+1) )
at x = dx ∗ (N − 1) ∗ 0.5, with N = 39. The results are shown in figure A.3.
Note that the Hankel function is designed as the inverse Fourier transform of the
phase shift operator and not as the Hankel function directly in the spatial domain.
In this way aliasing of the higher wavenumbers is avoided. It is observed that
the wavenumber spectrum is now unconditionally stable for all wavenumbers, but
the accuracy is reduced for higher wavenumbers. Another disadvantage is that the
effective wavenumber band, in which the filtered analytical operators work, is not
under control. It will be shown that control over the bandwidth of the spatial
operator is essential for a desired operator functionality. Another and better way
to obtain stable extrapolation operators is to start in the wavenumber domain and
transform the wavenumber expression back to the space domain. There are many
ways to do this transformation in a desired way. In the next subsection four different
transformation methods, including the new method, will be discussed.
A.1.2
From wavenumber domain to spatial convolution operators
The most straightforward way to obtain a spatial convolution operator from the
expression in the kx − ω domain is to calculate the inverse Fourier transform of
this expression in the space-frequency domain and truncate this operator to a
desired convolution length. Unfortunately this simple method is not stable for all
wavenumbers and can therefore not be used in a recursive migration scheme.
Appendix A: Operator optimization
189
1.5
1.0
amplitude
phase
1.0
0.5
0
-0.5
-1.0
amplitude
0
0.6
0.4
0.2
a.
-1.5
0.8
0.1
0.2
0.3
0.4
0.5
0
0.5
1.0
0.4
0.8
0.3
0.6
0.2
0.4
0.1
0.2
c.
0
0
b.
0
100
150
operator length [m]
200
250
0.2
0.1
0.2
0.3
0.4
0.5
0.3
0.4
0.5
d.
0
50
0.1
0
cycles
Figure A.4 Due to small derivatives of the operator in the wavenumber domain the artifacts for the
truncated operators remains small. The top pictures show the operator design in the
wavenumber domain and the lower pictures show the 39 point operator and its related
amplitude spectrum. Note the short spatial length of the spatial operator.
Smoothed wavenumber operator
Another way to obtain a useful operator is to design a kx − ω operator in such a way
that the inverse Fourier transform is short, so that truncation does not significantly
effect the spectrum of interest. The trick in this method is that the phase of the
phase shift operator, given by kz ∆z which has a singular point for kx = k, is altered
such that it is smoothly deviating from the correct phase for angles larger than
the maximum design angle. The extrapolation operator is then obtained by taking
the exponent of this smoothed phase, exp (−jφ(kx )). This function is then filtered
with respect to the maximum design angle to control the amplitude behavior of the
operator. In this way the amplitude spectrum of the operator is not decaying with
the exponentially slope known from the phase shift operator but is determined by
the slope of the used wavenumber filter. It is known from Fourier analysis that large
amplitude derivatives in the wavenumber domain as well as large phase derivatives
give rise to large operator lengths (Blacquière, 1989). So the wavenumber filter
should keep the operator as smooth as possible to arrive at a short spatial operator,
π
= kN yquist . The amplitude
for example by using a filter with a cosine taper to ∆x
spectrum of the operator is shown in figure A.4b for a maximum design angle of
70◦ . The phase of this operator is smoothly deviating from the exact operator (the
solid line in figure A.4a) for angles greater than the design angle. In figure A.4c
the truncated operator (39 points) and in figure A.4d the wavenumber spectrum is
190
A.1 1-Dimensional operators for 2-dimensional extrapolation
shown. In comparison with figure A.2 it is observed this truncated operator is stable
for all wavenumbers of interest.
For the methods discussed above the performance of the method was determined
from the wavenumber spectrum of the operator. The final aim in the design of a
spatial convolution operator is an optimal solution. The optimized result must be
designed accurate within the band of interest and stable outside this band. In order
to solve this problem the notion accuracy must be defined in a mathematical way.
Therefore a weighted error function, which measures the deviation form the true
function, is defined as
||ε(x)|| = || W (x)|Y (x) − P (x)| ||.
(A.6)
where Y (x) is the true function and P (x) the approximation. There are several
possible choices available for the norm function (k k) in equation (A.6). The most
widely used are the L∞ (also called the minimax, or Chebychev) norm and the L2
(least-squares) norm. The L2 and L∞ norm are respectively defined as
||Y (x)||2 ≡
Z
b
a
2
|Y (x)| dx
||Y (x)||∞ ≡ max |Y (x)|
a≤x≤b
12
(A.7)
(A.8)
Peaks and overshoots in the frequency domain are typical of frequency sampling
and least-squares designs. Windowing techniques, as discussed above, are attempts
to reduce the peaks in the error function (Parks and Burrus, 1987). The theory of
Chebychev approximation provides algorithms to find the coefficients of a convolution operator with the minimum value for the maximum error in the peaks. Operators that have this minimum value of the maximum error exhibit an equiripple
behavior in their frequency response. A practical reason for using the L∞ norm is
that when in computer calculations a complicated mathematical function is estimated by one that is easy to calculate then it is usually necessary to ensure that the
greatest value of the error function ε is less than a fixed amount. This is just the
required accuracy of the approximation which is a condition on the norm ||Y − P ||∞ .
After defining the accuracy in a desired way a computation method must be found
to calculate the solution of the approximation problem. By using the L∞ norm the
minimization problem is usually not solvable explicit in terms of formulas. However, by formulating the L∞ optimization problem as a Chebychev approximation
problem a set of conditions is provided which completely characterize the optimal
filter (Powell, 1981; Thorbecke, 1995). The Remez exchange algorithm, which is discussed briefly below, calculates the unique solution in the L∞ norm. The weighted
least-squares method which satisfies the L2 norm is discussed first.
Appendix A: Operator optimization
191
Weighted least-squares
The goal in the optimization procedure is to obtain a short spatial convolution
operator which has a wavenumber spectrum which is, over a desired wavenumber
band, equal or close to the exact formulation in the frequency-wavenumber domain.
This problem can be written as an integral equation given by
Ỹ (kx ) =
Zx2
x1
exp (jkx x)Y (x)dx for k1 ≤ kx ≤ k2 .
(A.9)
In this integral equation the integration is carried out over a limited spatial domain, representing the short operator, and the frequency-wavenumber domain of
the operator is bandlimited. The discrete counterpart of this integral equation is
Ỹ (n∆kx ) = ∆x
M2
X
m=M1
exp (jn∆kx m∆x)Y (m∆x) for N1 ≤ n ≤ N2 .
(A.10)
Written more explicitly in matrix notation
2
6
6
6
6
6
6
6
6
4
2
3
Ỹ (N1 ∆kx )
exp (jN1 ∆kx M1 ∆x)
6
7
.
.
6
7
.
.
6
7
.
.
6
7
6
7
1
Ỹ (0)
7 = ∆x 6
6
7
.
.
6
7
.
.
4
5
.
.
Ỹ (N2 ∆kx )
exp (jN2 ∆kx M1 ∆x)
...
..
.
...
...
...
1
.
.
.
1
.
.
.
1
...
...
...
..
.
...
3
32
Y (M1 ∆x)
exp (jN1 ∆kx M2 ∆x)
7
6
7
.
.
7
76
.
.
7
76
.
.
7
76
7
6
7
1
7 6 Y (0) 7
7
6
7
.
.
7
76
.
.
5
54
.
.
exp (jN2 ∆kx M2 ∆x)
Ỹ (M2 ∆x)
or
Ỹ = Y
(A.11)
with Y the desired short operator and Ỹ being its spatial Fourier transform, yielding an approximation of the exact phase shift operator. Further m = M1 . . . M2
represents the length of the desired short operator and n = N1 . . . N2 the length of
the Fourier transformation and where the wavenumber sampling is given by
∆kx =
2π
.
(N1 + N2 + 1)∆x
(A.12)
The number of samples in the frequency-wavenumber domain (N1 + N2 + 1) must
be chosen in such a way that the short spatial operator is zero outside its working
length. This means for the extrapolation operators that the number of samples must
be greater or equal to the number of traces to be extrapolated.
Matrix equation (A.11) has, with respect to the unknown spatial operator, more
equations than unknowns, meaning that it is usually impossible to find an unique
solution which satisfies all the equations. To solve this problem a solution is sought
which approximately satisfies all the equations in a least-squares manner. Therefore
the following error function is defined
ε̃ = Ẽ H Λ̃Ẽ,
(A.13)
192
A.1 1-Dimensional operators for 2-dimensional extrapolation
ΓH
Λ̃
Γ
Y
ΓH
Λ̃
Ỹ
Figure A.5 The matrices used in a weighted least-squares implementation of a Fourier inversion.
Note that if Λ̃ = I the matrix equation is equivalent to the DFT.
where superscript H denotes complex-conjugate transpose, with
Ẽ = ΓY − Ỹ
(A.14)
and Λ̃ a diagonal matrix containing a weighting function on its diagonal. In this
so called weighted least-square procedure (abbreviated as WLSQ) the weighting
function is defined in such a way that the wavenumbers of interest are given a
high weight. The wavenumbers which are not of interest are ‘abused’ to make the
operator as good as possible (given a low weight factor). The method described by
Hale (1991b) also uses some part of the wavenumber spectrum (called the degrees
of freedom) to force the amplitude spectrum to zero in the evanescent region. So
by introducing the weighting function a good control is obtained for the desired
function of the space-frequency operators. The least-squares solution of equation
(A.13) is given by
∂ε
= 0 ∀(Y H )i
(A.15)
∂(Y H )i
with
ε̃ = Y H ΓH − Ỹ H Λ ΓY − Ỹ
(A.16)
The solution of equation (A.16) is given by the normal equations (slightly modified
after Claerbout (1976))
ΓH Λ ΓY − Ỹ = 0
(A.17)
or
h
i−1
Y = ΓH Λ̃Γ
ΓH Λ̃Ỹ
(A.18)
In figure A.5 the weighted least-squares solution of equation (A.18) is given in a
matrix representation. The components of the Fourier transform and in the inverse
Fourier transform matrices are given by
Γnm = exp (jn∆kx m∆x)
ΓH
mn = exp (−jn∆kx m∆x)
amplitude
Appendix A: Operator optimization
0.5
1.0
0.4
0.8
0.3
0.6
0.2
0.4
0.1
0.2
0
193
0
0
50
100
150
operator length [m]
200
250
0
0.1
0.2
0.3
0.4
0.5
cycles
Figure A.6 Weighted least-square operator (left) of 39 points and its wavenumber spectrum (right).
The solid line is related to the spectrum of the exact phase shift operator and the dashed
line to the optimized operator. Note the perfect match in the propagating part of the
wavenumber domain.
The weight function, which is a diagonal matrix, is given by
Λnm = w(n∆kx )δnm
The total matrix, ΓH Λ̃Γ which stands before the unknown, is a square M × M
matrix which has to be inverted. For the 1-dimensional optimization problem this
matrix has a Toeplitz structure and can be inverted very fast by using the Levinson
scheme. If in equation (A.18) the weight matrix is chosen identical to the unit
matrix I then the right hand side of equation (A.18) is an inverse Fourier transform
(N-points) which is truncated to M-points in the spatial domain. In this specific
case no optimization is carried out.
For the extrapolation operator it is possible to neglect the operator behavior on
the evanescent waves because they are not present in the data above the noise
level. The recursive extrapolation scheme only demands a stable (amplitude < 1)
behavior of the wavenumbers in the evanescent region. The used weighting function
can therefore be a simple block with a weight of one inside the range of angles of
interest (the propagating waves) and a small value (1e-5) outside this band. How
these optimized operators, with the same parameters as before, behave is shown in
figure A.6. The spatial operator has a less ‘smooth’ character, but the wavenumber
spectrum is stable for all wavenumbers and is accurate within the band of interest.
Note the special character of the evanescent part of the wavenumber spectrum. It
is not smoothing down but varying. This is observed better in figure A.7 where
a collection of operators is shown, again with the same parameters as before but
within a frequency range from 0 to 60 Hz. Holberg (1988) used a same illustration
and argued that when the phase and amplitude errors are oscillating functions of the
wavenumber these errors will not accumulate at the maximum possible rate when
waves are propagated through inhomogeneous media. So the obtained operators
could be used in a stable manner in a recursive migration scheme. In the calculation
done to obtain figure A.7 a search for a suitable weighting function was carried
194
A.1 1-Dimensional operators for 2-dimensional extrapolation
Figure A.7 Collection of wavenumber-frequency spectra of the weighted least-squares operators
with the same parameters as before but with varying k = ωc . The evanescent region is
stable and the propagating region is accurate and stable for the whole k range.
out in such a way that the amplitude of the wavenumber operator is ensured to be
< 1.0001.
Note that the WLSQ method can also be used to obtain other spatial convolution
operators, for example elastic decomposition operators as described in Herrmann
(1992). The method can be further improved by using more optimization steps;
for example the Lawson algorithm (Rice and Usow, 1968), which will adjust the
weight function in such a way that after several steps the solution will converge
to a Chebychev-norm (L∞ ) solution see for example Algazi et al. (1986). In the
next subsection a spatial convolution operator with a Chebychev norm is obtained
directly, by using the Remez exchange algorithm.
Remez exchange algorithm
The Remez exchange algorithm is based on the L∞ norm and the related Chebychev
polynomial. The Chebychev polynomial is a powerful function in approximation
theory, because of the special properties of the Chebychev polynomials (Parks and
Burrus, 1987). The Chebychev approximation P (x) of a real function Y (x) is defined
as
Y (x) ≈ P (x) =
M
X
a0
+
am Tm (x)
2
m=1
(A.19)
for −1 ≤ x ≤ 1, with Tm (x) the Chebychev polynomial of order m and am the
expansion coefficients (see for example Johnson and Riess, 1977; Ralston, 1967; Kogbetliantz, 1960). The Chebychev polynomials Tm (x) are usually defined in terms
Appendix A: Operator optimization
195
of trigonometric functions by:
Tm (x) = cos (m arccos (x))
−1
Tm (x) = cosh (m cosh
(x))
for|x| ≤ 1
for|x| > 1.
Using the variable substitution x = cos (φ) the Chebychev polynomials can be rewritten as
Tm (x) = cos (mφ).
With this variable substitution some properties of the Chebychev polynomials can
easily be derived
2Tm (x)Tn (x) = Tn+m (x) + Tn−m (x)
Tm+1 (x) = cos ((m + 1)φ) = 2 cos (φ) cos (mφ) − cos ((m − 1)φ)
= 2xTm (x) − Tm−1 (x)
(A.20)
(A.21)
2
(A.22)
3
(A.23)
T2 (x) = 2x − 1
T3 (x) = 4x − 3x etc.
with T0 (x) = 1 and T1 (x) = x. The most important property is the recurrence
relation given in equation (A.21) and will be used in the derivation of 2-dimensional
convolution operators (used for 3-dimensional extrapolation) discussed in the next
P
section. Equations (A.22) and (A.23) show that Tm (x) can be written as
bm xm ,
a polynomial in x. The Chebychev polynomial have some special properties, for
example Tm (x) has m zeros which all fall in the interval [−1, 1] and are located at
the points
π(2i + 1)
)
(A.24)
xi = cos (
2m
The polynomials oscillate between +1 and -1 for −1 ≤ x ≤ 1 and go monotonically
to ±∞ outside that domain. In figure A.8 the Chebychev polynomials for the 0,1,2,3
and 4th order are plotted.
By using the orthogonal property of the Chebychev polynomials it can be proven
(Powell, 1981, p.144) that the Chebychev approximation P (x) in equation (A.19)
is very nearly the minimax polynomial, which (among all polynomials of the same
degree) has the smallest maximum deviation for the true function Y (x). A property
of this particular approximating polynomial of equation (A.19) is that it can be
truncated to a polynomial of lower degree n ≤ M that does again yield the ”most
accurate” approximation of degree n. So the accuracy of the approximation improves
when the number of terms is increased. Since the Tm (x) are all bounded between
±1, the difference between the truncated(n) and the larger polynomial(M) can be
no larger than the sum of the neglected terms. In fact if the terms are rapidly
decreasing, then the error is dominated by cn Tn (x) an oscillatory function with
196
A.1 1-Dimensional operators for 2-dimensional extrapolation
1.0
amplitude
0.5
0
-0.5
-1.0
-1.0
-0.5
0
0.5
1.0
Figure A.8 Chebychev polynomials for the 0,1,2,3 and 4th order within the interval [−1, 1]. Note
that outside this interval the polynomials of orders ≥ 2 approach ±∞ .
(n+1) equal extreme distributed smoothly over the interval [−1, 1]. This smooth
spreading out of the error is a very important property Press et al. (1992).
The Remez Exchange algorithm is to the Chebychev approximation as the normal
equations are to the minimum least-squares solution. In both cases a set of expansion
coefficients ai is calculated that ”best fit” a set of basis functions to the data. If
the definition of ”best fit” is to minimize the sum of squares then get the normal
equations are used to calculate the ai ’s. If the ”best fit” is defined to minimize the
maximum error (min-max or Chebychev fit) then the Remez Exchange algorithm
gives the answer. The Remez exchange algorithm for Chebychev approximation
converges from any initial reference to his unique answer.
The McCLellan and Parks (1972); McClellan et al. (1973) algorithm, which is used
in the extrapolation operator calculation, is based on the Remez exchange algorithm
and Chebychev approximation theory. The operators are optimal in the sense that
the maximum error between the desired frequency response and the actual frequency
response is minimized for a given weighting function. Operators designed this way
exhibit an equiripple behavior in their frequency response, and hence are sometimes
called equiripple filters. In figure A.9 the wavenumber spectrum of the spatial extrapolation operator, with the same parameters as before, is shown.
Non-linear optimization
The most advanced and complicated method to compute a spatial convolution operator is by non-linear optimization. Holberg (1988) and Blacquière (1989) have
both used a non-linear optimization method to compute stable and accurate extrapolation operators. In non-linear optimization methods an objective function
and a constraint function have to be defined. The objective function is defined in
that part of the wavenumber domain in which the operator must be accurate and
the constraint function is defined in the remaining part of the wavenumber domain
where the operator must be stable. The objective function used in the non-linear
amplitude
Appendix A: Operator optimization
0.5
1.0
0.4
0.8
0.3
0.6
0.2
0.4
0.1
0.2
0
197
0
0
50
100
150
operator length [m]
200
250
0
0.1
0.2
0.3
0.4
0.5
cycles
Figure A.9 Extrapolation operators obtained with the Remez exchange algorithm.
equiripple behavior in the evanescent part of the spectrum.
Note the
schemes is defined as the summation of the squared amplitude and phase errors over
all wavenumbers in the domain of interest. The constraint function is designed to
suppress the larger wavenumber to a value smaller than 1.0 in order to obtain stable
operators.
Two different non-linear optimization routines were used; the CFSQP routine written by Tits and Zhou (1996) and a (old) NAG lib routine. The initial guess, which
must be provided, is for both methods the truncated spatial operator. If the methods fails to find a solution with the provided initial guess the smoothed phase or the
weighted least-squares solution is taken as a new initial guess. The results for the
CFSQP algorithm are shown in figure A.10. Both non-linear methods give stable
and accurate results. A disadvantage of the NAG method is that the computation
time is very long and it does not always find a solution. However, a slight change in
the input parameters (for example by changing the maximum angle of interest) can
give a correct solution. Due to these practical disadvantages the CFSQP method is
prefered.
amplitude
In the foregoing subsections six types of optimization methods (Truncation, Gaussian
filtering, smoothed phase, weighted least-squares, Remez exchange and non-linear
0.5
1.0
0.4
0.8
0.3
0.6
0.2
0.4
0.1
0.2
0
0
0
50
100
150
operator length [m]
200
250
0
0.1
0.2
0.3
0.4
0.5
cycles
Figure A.10 Extrapolation operators obtained with non-linear optimization using the CFSQP algorithm.
198
A.1 1-Dimensional operators for 2-dimensional extrapolation
optimization) have been introduced for obtaining extrapolation operators. To make a
good comparison between the different methods their resulting phase and amplitude
spectra are plotted together into one figure. The same parameters are used as
before only this time with an operator length of 19 points to show the differences
more pronounced. The results are shown in figure A.11.
From the spatial wavelet in figure A.11 it is observed that the weighted least-squares
and the non-linear optimization procedure have relative high amplitude variations
for the greater offsets. So the contribution from the off-center points is greater in the
non-linear and the weighted least-squares methods than in the other methods. Application of these operators (with a long operator length) in strongly inhomogeneous
media can therefore give rise to artifacts in the extrapolation result. The other spatial wavelets decays more ‘smoothly’ from the off-center points. In the wavenumber
spectrum the differences between the operators can be interpreted more meaningful.
The phase and amplitude errors of the operators are displayed in the middle pictures
of figure A.11. From the wavenumber spectrum it is observed that the Gaussian operator is rapidly loosing amplitudes for higher angles. The truncated operator is
not stable for all wavenumbers especially for the wavenumbers just before the point
where the evanescent part starts (0.25 on the horizontal axis) the amplitude exceeds
amplitude 1 significantly. The smoothed phase operator is in this area more stable
and accurate than the truncated operator. The weighted least-square operator is
stable and accurate for all wavenumbers. In the evanescent region the amplitude is
larger than the other operators (but still smaller than 1). The phase errors are very
small and the amplitude errors are smaller than 0.001. The Remez exchange operator shows its nice equiripple behavior and is very accurate for both amplitude and
phase. The non-linear optimized operator is as good as the weighted least-squares
and Remez operator but has a little different behavior as shown in the bottom pictures of figure A.11. The non-linear solution has a relatively large error peak close to
the edge of the maximum design angle and may become unstable for waves traveling
with these wavenumbers.
In this subsection is has been shown that there are several ways to obtain a spatial
convolution operator. From the examples given above it can be concluded that the
spatial bandwidth of a desired spatial convolution operator must, in the optimization procedure, be constrained to a desired bandwidth. To overcome exponentially
amplifying of certain propagating wavenumbers in a recursive depth migration it
must also have a stable behavior. In the next subsection the discussed operators are
used in an explicit recursive migration scheme to observe the effects of the different
operators in a depth section. From the foregoing results it is not possible to say
something about the performance in a migration algorithm. It is interesting to see
how the differences observed in figure A.11 above can be found back in a migrated
depth section.
Appendix A: Operator optimization
0.5
amplitude
0.8
➌
0.3
➎
0.1
➋ ➊
0.4
➏
➍
➍
0
0
25
50
75
operator length [m]
100
0
0.1
0.2
0.3
0.5
0.10
➊
0.04
0.05
0
➊
0
➌
➌
➋
-0.04
➋
-0.05
0
0.05
0.10
0.15
cycles
0.20
0.01
-0.10
0.25
0
0.05
0.10
0.15
cycles
0.005
➎
0
0
0.05
0.20
0.25
➏
➏
phase
0.4
cycles
0.08
-0.01
➎
➏
0.2
0
phase
➌
➋
0.6
0.2
-0.08
➊
1.0
0.4
199
➎
➍
0.10
0.15
cycles
➊ truncated Fourier
➋ Gaussian taper
➌ smoothed phase
0.20
0
-0.005
0.25
0
➍
0.05
0.10
0.15
cycles
0.20
0.25
➍ Weighted least-squares
➎ Remez exchange
➏ non-linear cfsqp
Figure A.11 The two top pictures show the amplitude of six different 19 point extrapolation operators in space (left) and the related wavenumber spectrum (right). The middle pictures
show the phase (left) and amplitude (right) errors for the operators shown on the top.
The two bottom pictures show a detailed view in the propagating region for the phase
(left) and amplitude (right) errors.
A.1.3
Recursive depth migration
In recursive depth migration lateral varying convolution operators are used to extrapolate seismic data through inhomogeneous media. The extrapolation operators
in the x − ω domain are locally homogeneous operators for each grid point (x, z) as
show in figure A.12. Based on the frequency ω and the local velocity c in gridpoint
(x, z) the extrapolation operator is computed, or in general read from an operator
table which is computed in advance based on the frequency range of interest and
the minimum and maximum velocity found in the macro model. Note that the
200
A.1 1-Dimensional operators for 2-dimensional extrapolation
∆z
operator length
∆z
∆z
∆z
Figure A.12 The principle of recursive extrapolation with spatial convolution operators. In principle at every lateral position a new operator is used to extrapolate the data from one
depth level to another depth level. Within the operator length the medium is assumed
to be homogeneous.
homogeneous extrapolation operator is symmetric around its center and can be implemented in an efficient way. Because of the assumption of locally homogeneous
media, the extrapolation depth step should be small and the operator length short.
To obtain reliable extrapolation results the operators must be accurate and remain
stable in the recursion scheme. In the previous section it was shown that with these
constraints it is possible to use different methods to design the operators. In this
section six different extrapolation operators are considered: the truncated, smoothed
phase, Gaussian tapered, non-linear (CFSQP), Remez exchange, and the weighted
least-squares operator.
Homogeneous shot record migration experiments are carried out in a medium with a
velocity of 2000 m/s, a length of 2000 m and a depth of 1000 m with the spatial and
depth intervals, ∆x = ∆z = 10 m. The zero offset trace in the shot record contains
three Ricker wavelets at 0.3 , 0.6 and at 0.9 s (all the other traces are filled with
zeros). The source wavelet is sampled with 4 ms and has an amplitude spectrum
up to 60 Hz. In figure A.13 twelve depth sections are shown for six different extrapolation operators and two different operator lengths. The left-hand side pictures
give depth sections made with an operator length of 19 points and designed to be
accurate for angles up to 65 degrees. The right-hand side pictures represent depth
sections obtained with an operator length of 39 points and a maximum design angle
of 85 degrees. Note that the events observed in the depth images lie on concentric
semi circles with centers at the origin.
Appendix A: Operator optimization
201
The truncated operator results in figure A.13 is not stable for the 19 point operator.
In the deepest event in the depth section of the 19 point result the cumulative error
can be observed in the distorted semi circle. The 39 point operator is stable but the
edge effects of the operator disturb the image strongly. The migration results with
the Gaussian tapered operator has the smallest artifacts in it but the price is that
the higher angles are strongly attenuated and for the 19 point operator the deepest
event is almost vanished. The results obtained with the smoothed phase operators is
stable but contains noise for the 85 degrees experiment and the higher angles in the
65 degrees experiment are less accurate than the weighted least-square optimized
result. The weight function used in the WLSQ method to obtain these results is
the simple box function described earlier. Only the higher angles (> 65◦ ) in the 19
point operator are not perfect and the steeper dips are attenuated, which is due to
the short length of the operator, the 39 point operator gives almost perfect results.
The Remez exchange operators give also good results which are comparable with
the results obtained with the weighted least-square method. The result obtained
by using non-linear optimization operators shows for the 19 point operator a good
section up to the higher angles, at the higher angles there are some artifacts visible.
The non-linear 39 point operator gives a better result but it takes a very long time,
compared with the other methods, to compute the operator table.
In table A.1 the computation times of the operator table used for the different
methods of the migration experiments in figure A.13 are given. The operator table
which is calculated in advance consists of 101 operators. The truncated, Gaussian
and smoothed phase operator calculation use the same amount of computation time,
the weighted least-squares and the Remez exchange method consume more time
(respectively 3 and 10 times more). The non-linear optimization method takes a
very long time to compute all operators. It is therefore not very efficient to compute
the operator table with a non-linear optimization routines.
Method
19 points
39 points
Truncated
Gaussian
Smoothed phase
WLSQ
Remez
Non-linear
0.13
0.13
0.13
0.23
1.0
125.0
0.13
0.13
0.13
0.30
3.37
565.0
Table A.1 Computation time in seconds, on a DEC Alpha (3000-500), for the operator table used to
calculate the results in figure A.13.
The possibilities and limitations of the weighted least-squares procedure are illustrated with some simple examples. From a computational point of view the desired
202
A.1 1-Dimensional operators for 2-dimensional extrapolation
➊ 19 points
➊ 39 points
➋ 19 points
➋ 39 points
➌ 19 points
➌ 39 points
➍ 19 points
➍ 39 points
➎ 19 points
➎ 39 points
➏ 19 points
➏ 39 points
Figure A.13 Impulse responses for six different extrapolation operators. The left-hand side pictures
gives a depth section calculated with an extrapolation operator length of 19 points
which is designed to be accurate for angles up to 65◦ , the right-hand side pictures
represent a depth section obtained with an operator length of 39 points and a maximum
design angle of 85◦ .
Appendix A: Operator optimization
203
operator length must be made as short as possible. To illustrate the influence of
reducing the number of operator points the experiment of figure A.13 is repeated
for different operator lengths. The results are shown in figure A.14. The results obtained with the 11 point operator is inaccurate and unstable for the steeper dips, the
13 point operator is stable but for steeper dips it is still inaccurate. This inaccuracy
can be observed most clearly in the smallest semi circle. It is remarkable that the 13
point operator is so much better than the 11 point operator. The 15 point operator
is both accurate and stable, the 19 point operator is stable and even more accurate
but this is hardly visible. From these migration results it is difficult to predict how
an operator with a certain length behaves. The only practical rule is that a longer
operator in a homogeneous medium will always be more accurate and stable than
a shorter operator. The wavenumber spectra, for a representing frequency, of the
different operators gives more information about the accuracy and stability of the
operator. It is remarkable that the unstable 11 points operator has a stable behavior in the propagating region and is unstable for the evanescent region. This is
also observed in the migrated section which contains high frequency artifacts. This
behavior can be influenced by choosing a higher weight value for the wavenumber in
the evanescent region. However, this will also give rise to higher peaks, meaning less
accurate, in the propagating region. In the wavenumber spectra it is also observed
that by using more operator points in the spatial domain will give smaller amplitude
errors in the propagating region.
When the maximum angle of interest is reduced the operator length can also be
reduced to obtain the same result with a shorter computation time. In the limit a 1
point operator is sufficient to extrapolate a plane wave. In figure A.15 the wavenumber and phase spectrum is shown for a 7 point operator for different maximum design
angles. The parameters are the same as before, the frequency shown in the bottom
pictures of figure A.15 is 20 Hz. In the amplitude spectrum an interesting property
is observed; for higher design angles the operator has a bigger amplitude at the evanescent region, for the 50◦ operator even bigger than 1. This property is also observed
in the right-hand side depth section of figure A.15, this depth section, with operators
designed for 45◦ , is inaccurate for the higher angles. For the operator designed with
a maximum angle of 30◦ the depth section is more accurate and the higher angles
of propagation are attenuated. With these simple experiments it is shown that an
operator cannot be both short and accurate for steep dips. In designing a short
operator the higher angles must be attenuated to overcome steep dip distortion and
unstable properties.
In conclusion;
the weighted least-squares method is a good candidate to compute the operator table used in recursive depth migration schemes, because of its
stable and accurate behavior for a broad range of operators, and the relatively small
computation time needed to compute the operators. Reducing the operator length
will decrease the computation time, but for very short operators the steeper dips
204
A.2 2-Dimensional operators for 3-dimensional extrapolation
➍ 11 points
➍ 13 points
➍ 15 points
➍ 19 points
0.005
11
1.4
amplitude
1.2
15
1.0
11
0.8
0
0.6
13
13
0.4
19
0.2
0
15
0
0.1
0.2
0.3
cycles
19
0.4
-0.005
0.5
0
0.05
0.10
0.15
cycles
0.20
0.25
Figure A.14 Depth sections obtained with different operator lengths in the WLSQ method, but with
a fixed maximum design angle of 65◦ . The two bottom pictures show a detailed view of
the operator errors at 20 Hz. The artifacts which are observed in the depth section can
be found back in the behavior of the operator.
will become inaccurate and unstable. The weight function used in the optimization
procedure can also be optimized in order to give the spatial operator an even more
controlled behavior. This optimization of the weight function has been carried out
by Rice and Usow (1968) and is called the Lawson algorithm. Note that the length
of the optimized operator is for a certain error a function of the wavenumber. In
calculating the table of operators this length dependency could be taken into account
to arrive at an even more efficient extrapolation.
A.2
2-Dimensional operators for 3-dimensional extrapolation
To visualize the 3-D subsurface of the earth 3-D migration algorithms are needed
which give accurate results within a reasonable computation time. In this section
several recursive depth migration algorithms are discussed and compared with each
other. The backbone of every recursive depth migration method is a 3-dimensional
extrapolation algorithm. In lateral homogeneous media the extrapolation algorithm
Appendix A: Operator optimization
➍ 7 point, 30◦
➍ 7 point, 45◦
1.5
1.0
0.8
0.6
1.0
50
phase
amplitude
40
30
0.4
45
0.5
40
0
-0.5
-1.0
0.2
0
205
30 45
-1.5
0
0.1
0.2
0.3
0.4
0.5
0
0.1
0.2
cycles
0.3
0.4
50
0.5
cycles
Figure A.15 Using the WLSQ method with a fixed length of 7 points, the influence of different maximum design angle is investigated. Due to the behavior of the operator in the evanescent region (bottom pictures) high frequency artifacts are observed in the depth section.
can be a simple multiplication in the wavenumber domain, but extrapolation through
3-dimensional inhomogeneous media is a more computation intensive operation and
requires a space-variant 2-dimensional spatial convolution. Recently various authors
(Hale (1991b); Sollid and Arntsen (1994); Gaiser (1994); Biondi and Palacharla
(1994); Kao et al. (1994); Thorbecke and Berkhout (1994, 1995); Soubaras (1996))
have published articles which pay attention to an optimized calculation and efficient
implementation of 3-dimensional extrapolation operators in a recursive depth migration. This section will give an overview of the existing methods and introduces
several efficient optimization and implementation methods that have not yet been
discussed in the Geophysical literature. The computation times of the different algorithms are compared with each other and the performance of the extrapolation
algorithm is checked with the aid of a simple synthetic experiment.
A first subdivision between the different optimization methods can be made with
respect to the type of expansion of the analytical phase shift operator (equation
(A.1)) in the wavenumber domain. This expansion can be a (Taylor)
q series expansion
of the analytical phase shift operator with respect to kz =
k 2 − (kx2 + ky2 ) (as
presented in this section), an expansion with respect to kr2 = kx2 + ky2 (Berkhout
(1982), Soubaras (1996), Sollid and Arntsen (1994), Hoff (1995) and this section)
or an expansion with respect to the cosine terms of the 1-D Fourier transformation
Hale (1991a). In equations (A.27) to (A.29) these different approximation to the
206
A.2 2-Dimensional operators for 3-dimensional extrapolation
expansion
cos (kq
x ) cos (ky )
cos ( kx2 + ky2 )
q
k 2 − (kx2 + ky2 )
optimization
Least Squares
Weighted Least Squares
Non-Linear
Remez Exchange
kx2 + ky2
implementation
2-D convolution
1-D convolution
Chebychev
Series
Table A.2 Three criteria which are used to discriminate between different 3-dimensional extrapolation algorithms. Note that in principle many combinations between the elements in the
three blocks are possible.
phase shift operator are shown
Ỹ0 (kx , ky ) = exp (−jkz ∆z),
≈
≈
≈
≈
M
X
N
X
Ymn cos (kx m∆x) cos (ky n∆y),
(A.25)
(A.26)
m=0 n=0
M
X
Ym cos (kr m∆x),
(A.27)
am [kx2 + ky2 ]m ,
(A.28)
bm [kz ]m .
(A.29)
m=0
M
X
m=0
M
X
m=0
Equation (A.26) is the so called direct method which calculates the 2-D extrapolation operator Ymn by a direct optimization method and can be regarded as a
weighted expansion in 2-dimensional cosine terms. The operators obtained with
the direct method require the implementation of a full 2-dimensional spatial convolution. Equation (A.27) is a reduction of the 2-dimensional filter problem to a
1-dimensional filter problem by using the circular symmetry of the 2-D phase shift
operator and is represented by an expansion in 1-dimensional cosine terms. The
1-dimensional filter problem, to obtain Ym , can be solved with any preferred 1-D
optimization method. The cosine terms in equation (A.27) are approximated by
small 2-dimensional convolution filters.
Equation (A.28) and (A.29) are expansions of the operator in non-spectral polynomials. The terms am and bm in the series expansions can be obtained by calculating
the coefficients from a Taylor series expansion Berkhout (1982), or optimizing the
coefficients with an error definition in a preferred norm, for example the L2 or L∞
norm. The basic polynomials kx2 + ky2 and kz which occur in equations (A.28) and
(A.29) are approximated by short and accurate convolution operators.
A second subdivision between the different extrapolation algorithms can be made
Appendix A: Operator optimization
207
with respect to the kind of optimization method used to obtain the spatial convolution operator. The type of implementation of the spatial convolution is a third
criterion to discriminate between the different methods. table A.2 gives an overview
of the different techniques which can be used in the expansion, optimization and
implementation. Note that in principle many combinations between the elements
in the three blocks are possible. For example Holberg (1988) and Blacquière et al.
(1989) use a non-linear optimization technique for the operator optimization in a
2-dimensional cosine series (equivalent to a weighted inverse Fourier transformation
according to equation (A.26)) and have implemented this operator as a full 2-D convolution. Hale (1991a) makes use of the McClellan transformation in equation (A.27)
and uses the coefficients of a 1-D convolution operators in a Chebychev recursion
scheme. Soubaras (1996) uses the Remez exchange algorithm in the optimization
of the 1-D convolution operators and in the optimization of the expansion factors
(with respect to powers of kx2 + ky2 ) of the phase shift operator in equation (A.28). In
this section the weighted least squares optimization method is introduced as a fast
alternative method for the optimization of the 2-D convolution operators and in the
optimization of the factors in the series expansions. The McClellan method is discussed in detail and several schemes which optimize the original McClellan method
are discussed briefly. The series expansions with respect to kz and kx2 + ky2 given in
equation (A.28) and (A.29) are worked out in more detail and compared with the
other extrapolation methods.
A.2.1
Direct method
The most straightforward method which does not make use of any series expansion is called, after Berkhout (1982), the direct method. In the direct method the
optimization for the convolution operator is defined by the Fourier transformation
from wavenumber to spatial domain and implemented by means of an optimization
scheme (see Holberg (1988); Blacquière et al. (1989) for the results with a non-linear
optimization scheme). The advantage of the direct method is the uncomplicated optimization of the operator and the simple implementation. A disadvantage of this
direct method is that in the space-frequency domain the full 2-D convolution has
to be carried out for every spatial position. For an operator with a typical operator size of 25 × 25 points this means 625 complex multiplications and summations
for every grid point! By using the even symmetry in the operator the number of
multiplications can be reduced by a factor of 4 by folding the data into a quarter
and application of the convolution to this folded part. However the number of flops
remains high, especially if one takes into account that this convolution has to be
carried out for every gridpoint, for every frequency of interest and for all the depth
steps. In this section the results obtained with the direct method are used as a
reference result for comparison with the other, non-direct, methods. In this section
three different methods, to calculate the 2-D direct convolution operators, are dis-
208
A.2 2-Dimensional operators for 3-dimensional extrapolation
cussed with respect to accuracy and efficiency. To compare the results a synthetic
migration experiment is introduced which calculates the impulse response of the
used operators.
Weighted Least Squares optimization
The most simple way to obtain space-frequency operators is an inverse Fourier transformation of the exact operators from the wavenumber-frequency domain back to
the space-frequency domain. Despite of its simple form this solution is not very
efficient because the spatial convolution operator obtained in this way must be very
long to give stable and accurate results. Tapering the spatial operator gives some
improvements (Nautiyal et al., 1993) but for accurate extrapolation results tapering
cannot be used as pointed out by in the previous section. The aim in the design of
the operator is a short convolution operator with a wavenumber-frequency spectrum
which is, over a desired wavenumber band, equal or close to the phase shift operator in the kx , ky -ω domain. The starting point in the analysis of this optimization
problem is the inverse Fourier transformation which is defined as
Ỹ (kx , ky ) =
Z Z∞
Y (x, y) exp (jkx x) exp (jky y)dxdy.
(A.30)
−∞
Using the discrete version of the Fourier integral and the even symmetry in the phase
shift operator equation (A.30) is rewritten in a discrete equation (Blacquière, 1989,
after)
Ỹ (kx , ky ) ≈ ∆x∆y
M X
N
X
Smn Y (m∆x, n∆y) cos (kx m∆x) cos (ky n∆y),
(A.31)
m=0 n=0
with Smn defined as
Smn =



1
2


4
for m = n = 0,
for m = 0 ∨ n = 0, .
(A.32)
for n 6= 0 ∧ m 6= 0
Using the Circular symmetry in the operator by interchanging n and m and the fact
P
Pm
PN PM
that M
m=0
n=0 =
n=0
m=n the number of equations can be further reduced to
1/8 of the original number of equations (this reduction is only possible if ∆x = ∆y)
Ỹ (kx , ky ) ≈
+
∆x2
M X
m
X
m=0 n=0
h
Y (m∆x, n∆y) Smn cos (kx m∆x) cos (ky n∆y)
i
,
′
Snm
cos (kx n∆x) cos (ky m∆y)
(A.33)
Appendix A: Operator optimization
′
with Snm
defined as



0 for n = m,
2 for n = 0 ∨ m = 0,


4 for m =
6 0 ∧ n 6= 0
209
(A.34)
and M × N being the user specified size of the desired short operator. Using the
shorter matrix notation equation (A.33) can be rewritten as
Ỹ ′ = Y ,
(A.35)
with Ỹ the desired short operator and Ỹ ′ being its spatial Fourier transform, yielding an approximation of the exact phase shift operator. Equation (A.31) with the
quarter, or equation (A.33) with the octal part has to be solved for the unknown
operator coefficients Ymn = Y (m∆x, n∆y) for all wavenumbers (kx , ky ) of interest.
Therefore the same weighted error function ε̃ as introduced in the previous section
is used and the same solution is obtained
h
i−1
Y = ΓH Λ̃Γ
ΓH Λ̃Ỹ ,
(A.36)
where ΓH Λ̃Γ is a square matrix. The weighted least-squares method can be used in
the calculation of short (2-D) spatial convolution operators but also in the calculation of series expansion factors discussed later. For the 1-dimensional optimization
problem the WLSQ method can be inverted fast by using the Levinson scheme.
For the 2-dimensional problem, the Toeplitz structure is not present anymore and
standard LINPACK routines are used to calculate a QR decomposition of the matrix ΓH Λ̃Γ and with this decomposition the solution of matrix equation (A.36)
can be found. In figure A.16 the wavenumber spectrum of a WLSQ optimized,
19 × 19 points spatial convolution operator is given for 128 × 128 kx , ky points with
c = 1000 m/s, f = 25 Hz, ∆x = ∆y = ∆z = 10 m and a maximum angle of interest
(αmax ) of 65◦ . In the remainder of this section, about 2-dimensional operators,
these parameters will be used in all further examples which represent a phase shift
operator in the wavenumber domain. The WLSQ method gives an accurate operator
which has a wavenumber spectrum close to the exact phase shift operator as shown
in figure A.16. Note that due to the optimization on a rectangular grid the operator
has a somewhat square symmetrical shape. If ∆x = ∆y the matrix definition which
uses the octal symmetry, given in equation (A.33), can be used in the implementation of the WLSQ solution. This scheme reduces the operator calculation time with
a factor of 2 in comparison with equation (A.31), where a quarter of the operator is
taken into account, and the matrix equation contains less degrees of freedom so the
unknown parameters are better defined.
Impulse response of an extrapolation operator
An impulse response experiment is used to test the behavior of the extrapolation
operator in an explicit recursive depth migration algorithm. In the middle of a spatial
210
A.2 2-Dimensional operators for 3-dimensional extrapolation
Figure A.16 The wavenumber spectrum of a WLSQ optimized operator with 19 × 19 spatial points,
128 × 128 kx , ky points with c = 1000 m/s, f = 25 Hz, ∆x = ∆y = ∆z = 10 m
and a maximum angle of interest of 65◦ .
limited homogeneous medium a point source is defined with the source signature
shown in figure A.17. This zero-phase Ricker wavelet is centered at 0.512 seconds
and has its frequency peak at 13 Hz. The constructed input data set, which contains
the wavelet, is transformed to the frequency domain and extrapolated to deeper
depth levels for every frequency of interest. At every depth level an imaging step is
carried out and the depth image is stored in memory. At the end of the calculation
for all frequencies the final depth image is written to disk. The block-diagram of this
algorithm is shown in figure A.18. For the other extrapolation algorithms discussed
in this subsection the operator optimization block in figure A.18 is replaced with the
optimization method of interest, everything else in the scheme remains the same. For
this synthetic experiment the following parameters are used: c = 1000 m/s, fmin =
1 Hz and fmax = 45 Hz, ∆x = ∆y = ∆z = 10 m, ∆t = 0.004 s and 55 depth steps
are taken on a x,y grid of 111×111 samples wide. Note that the maximum frequency
π
(= kN ).
is positioned in the wavenumber domain at 0.9 ∗ ∆x
9
amplitude
amplitude
1.0
0.5
0
-0.5
0
0.2
0.4
0.6
time [s]
0.8
1.0
6
3
0
0
10
20
30
40
frequency [Hz]
50
60
Figure A.17 Time signature (left) and Amplitude spectrum (right) of the wavelet used in the migration experiments.
Appendix A: Operator optimization
211
input parameters
initializations
operator calculation
next
frequency
extrapolation x and y
next
depth
imaging
writing depth image
Figure A.18 Processing scheme for the impulse response experiment. Note that the extrapolation
and the operator calculation block in the scheme is different for every different extrapolation method discussed in this section.
The exact response for the described experiment can be calculated by making use
of the dipole pulse response which is given by
G0 (r, k, φ) =
1 + jkr
cos φ exp (−jkr),
r2
(A.37)
p
with k = ωc , cos φ = zr and r = z 2 + x2 + y 2 . Using the complex conjugate G∗0 of
the dipole response in a non-recursive version of the scheme given in figure A.18 a
reference impulse response can be calculated for the synthetic experiment described
above. This reference response is discussed first to indicate what the interesting
points in the impulse response are.
In figure A.19 the time responses for several depth steps are shown together in one
picture. The imaging step at a certain depth level is equivalent with selecting the
zero time value for all x- and y-positions. The lowest event in figure A.19 is the time
response of the pulse after an inverse extrapolation step of 100 m, every higher event
represents a depth level 100 m deeper. It is observed that for the deeper event the
crossing with t = 0 is converging to the x = 0 position and will finally disappear if
the depth exceeds 512 m, it is also observed that every depth slice corresponds to
a certain dip angle. For example for a depth slice at 200 m the dip angle is given
−1
by cos φ = z(ct0 ) ⇒ φ ≈ 67◦ . In figure A.20 three cross-sections out of the 3-D
depth image of the reference experiment are shown; the right-hand side pictures
212
A.2 2-Dimensional operators for 3-dimensional extrapolation
-500
-250
0
250
500
0
0.1
0.2
0.3
0.4
0.5
Figure A.19 Time responses for several depth steps. The lowest event is the time response of the
pulse after an inverse extrapolation of 100 m every higher event represents a depth
level 100 m deeper.
in figure A.20 show a vertical cross-section for x = 0 (top picture) and a diagonal
section for x = y (bottom picture) and the left hand side picture in figure A.20
shows a horizontal cross-section at a depth of 220 m which corresponds to a reflector
dip of 65 degrees. Note that everywhere in this section the presentation of impulse
responses will be the same as presented in figure A.20, meaning that the clipping
level of the gray scales is the same for all the shown impulse response.
√ Note also
that in the diagonal cross-section the distance between two traces is 2∆x, where
∆x is the distance between the traces in the vertical x = 0 section.
Using the 2-dimensional 19 × 19 convolution operators obtained with the introduced
WLSQ method gives the depth image shown in figure A.21a. In the spatial convolution scheme the even symmetry in the operator is used explicitly by folding the
data into common operator point parts, which reduces the number of multiplications significantly. In the calculation of the convolution operator only 1/8 th of
the total spectrum is used by making use of the circular symmetry in the operator
and the fact that ∆x = ∆y. In figure A.21a it is observed that the artifacts in
the depth image consists of inner ’circular’ events at the higher angles which have
a square structure. This square structure is due to the fact that the solution of
the optimization problem is calculated on a rectangular grid. In the presentation of
the paper of Kao et al. (1994) similar features were observed. Using a longer 2-D
convolution operator as shown in figure A.21b for a 25 × 25 points operator these
rectangular artifacts have almost vanished. A more detailed discussion of the errors
in the extrapolation operators will be given below.
Appendix A: Operator optimization
-500
-250
x-position [m]
0
250
-500
500
-250
213
0
250
500
0
-500
250
y-position [m]
-250
0
500
0
250
250
500
z = 220
500
x=0
x=y
impulse response for reference operator
Figure A.20 Reference output for the migration experiment with left the depth slice at z = 220 m.
Top right shows a vertical slice for x=0 and bottom right a diagonal slice for x=y. Note
the perfect circular shape and the accuracy at the higher angles.
Hankel Transformation
A disadvantage of the direct optimization method discussed in the previous section is
that the use of a rectangular grid is visible in the results and the circular symmetry
of the operator is not used to its limits. By using the circular symmetry in the
phase shift operator the operator optimization problem is better defined because
less equations are used to solve for the unknowns. This may reduce the artifacts
caused by the use of a rectangular grid and will consume less computation time. The
circular optimization problem can be derived by rewriting the continuous Fourier
transform pair
Z Z ∞
Ŷ (kx , ky ) =
Y (x, y) exp (jkx x) exp (jky y)dxdy,
−∞
Z Z ∞
1
Y (x, y) = 2
Ŷ (kx , ky ) exp (−jkx x) exp (−jky y)dkx dky ,
4π
−∞
with the aid of polar coordinates.
For a circular symmetric function in the wavenumq
2
2
ber domain with kr = kx + ky , kx = kr cos θ, ky = kr sin θ and the Jacobian
dkx dky = kr dkr dθ the continuous inverse Fourier transform can be rewritten to
1
Y (x, y) =
4π 2
Y (x, y) =
1
4π 2
Z
2π
0
Z
0
Z
∞
Ŷ (kr ) exp (−jkr x cos θ) exp (−jkr y sin θ)kr dθdkr ,
0
∞
Ŷ (kr )kr dkr
Z
0
2π
exp (−jkr (x cos θ + y sin θ))dθ.
214
A.2 2-Dimensional operators for 3-dimensional extrapolation
-500
-250
x-position [m]
0
250
-500
500
-250
0
250
500
-250
0
250
500
0
-500
250
y-position [m]
-250
0
500
0
250
250
500
z = 220
500
x=0
x=y
a. 19x19 point convolution operator
-500
-250
x-position [m]
0
250
-500
500
0
-500
250
y-position [m]
-250
0
500
0
250
250
500
z = 220
500
x=0
x=y
b. 25x25 point convolution operator
Figure A.21 Results of the WLSQ optimized operators for different operator sizes. Note that the
artifacts which are present in the result for the 19x19 operator disappear for the larger
operator size.
Introducing r = x cos ϕ + y sin ϕ =
Y (r, ϕ) =
1
2π
Z
0
∞
p
x2 + y 2 (x = r cos ϕ, y = r sin ϕ) gives
Ŷ (kr )kr dkr
1
2π
Z
0
2π
exp (−jkr r cos (θ − ϕ))dθ,
(A.38)
Appendix A: Operator optimization
215
Figure A.22 The wavenumber spectrum of a circular WLSQ optimized operator with 19×19 spatial
points and 128 × 128 kx , ky points with c = 1000 m/s, f = 25 Hz, ∆x = ∆y =
∆z = 10 m and a maximum angle of interest of 65◦ . This operator is not useful in a
recursive extrapolation scheme.
with the definition of the zero order Bessel function as (equation 9.1.18 in Abramowitz and Stegun (1968) )
Z
1 π
cos (τ cos (θ))dθ,
J0 (τ ) =
π 0
Z π
1
J0 (τ ) =
exp (−jτ cos (θ)) + exp (jτ cos (θ))dθ,
2π 0
Z 2π
1
exp (−jτ cos (θ))dθ
J0 (τ ) =
2π 0
substituted into equation (A.38) gives
Z ∞
1
Y (r) =
Ŷ (kr )J0 (kr r)kr dkr .
2π 0
(A.39)
This equation expresses that the spatial convolution operator and its Fourier transformation are both circular symmetric and are related by the Hankel transform.
This is illustrated by the following formulations in polar coordinates
Z ∞
1
Y (r) =
Ŷ (kr )J0 (kr r)kr dkr ,
2π 0
Z ∞
Ŷ (kr ) = 2π
Y (r)J0 (kr r)rdr.
0
The spectral limited and discrete version of the Hankel transform is given by
Y (p∆r) =
N
1 X
Ŷ (n∆kr )J0 (n∆kr p∆r)n∆kr ,
N ∆r n=0
216
A.2 2-Dimensional operators for 3-dimensional extrapolation
Figure A.23 The wavenumber spectrum of a rotated Fourier reconstructed operator with 19 × 19
spatial point, 128 × 128 kx , ky points with c = 1000 m/s, f = 25 Hz, ∆x = ∆y =
∆z = 10 m and a maximum angle of interest of 65◦ .
with ∆kr = N2π
∆r . This last equation is implemented in the WLSQ optimization
scheme of equation (A.36) in which the matrix Γ is defined by the zero order Bessel
function J0 (kr r) in place of the cosine terms of the Fourier transform. The solution
of this problem is an optimized short operator as function of r. In the optimization
problem we can choose the points r in such a way that they coincide with the
rectangular spatial grid the extrapolation is carried out. The wavenumber spectrum
of the calculated solution is shown in figure A.22, whith the same parameters as
used in figure A.16. The spectrum shown is far from good and cannot be used in
an extrapolation algorithm. The problem with the Hankel transformation is that
for the spatial position r = 0 it is not possible to define a suitable value. If we
make it zero we get a singular matrix and making it non-zero is a random choice.
So the Hankel transformation cannot be used directly to design circular symmetric
convolution operators.
Rotated Fourier reconstruction
The idea of rotating a 1-dimensional operator is very useful in the wavenumber
domain. In the previous section it was shown that the desired 2-dimensional convolution operator must have a circular symmetric frequency response. The projection
onto a line oriented with an angle φ from one of the spatial axes is identical with
an optimal 1-D convolution operator (Kato and Matsumoto, 1982). A slice along
the same orientation in the Fourier domain corresponds to the Fourier transform of
the 1-D convolution operator. This is known as the projection slice theorem. Thus
the circularly symmetric frequency response is exactly described by one single projection. The problem of obtaining the circular 2-D convolution operator from the
spectrum of the optimized 1-D convolution operator can be solved is several ways.
Appendix A: Operator optimization
-500
-250
x-position [m]
0
250
-500
500
-250
217
0
250
500
0
-500
250
y-position [m]
-250
0
500
0
250
250
500
z = 220
500
x=0
x=y
19x19 point convolution operator
Figure A.24 Results of the rotated Fourier reconstructed operators for a 19x19 point convolution
operator. Note that the artifacts in the result for the 19x19 operator will become weaker
by using a larger operator.
The McClellan transformation is one of them and described in subsection A.2.2.
Another method is the Fourier reconstruction method which is described below.
Basically the rotated Fourier reconstruction method consists of designing a 1-D
optimized spatial operator, computing and rotating its frequency response in the 2D wavenumber plane, filling the undetermined region with the Nyquist value of the
1-D wavenumber response, performing the 2-D inverse Fourier transformation and
then windowing the result. The outer region of the circle is filled with the Nyquist
value of the 1-D spectrum to eliminate the Gibbs phenomenon. Note that the
rotated wavenumber spectrum has the circular symmetry only in the circular centum,
therefore the projection of the 2-D spectrum is only identical with the 1-D spectrum
along the vertical and horizontal directions. The projection along any other direction
is not exactly identical with the 1-D spectrum but gives a good approximation.
After the inverse Fourier transform the result is truncated to the original length of
the 1-D convolution operator (NxN). Since the projection is of finite length N, the
obtained 2-dimensional convolution operator has a nearly finite support. Therefore
the rectangular windowing distorts the circular frequency response only slightly.
The wavenumber spectrum of an operator obtained in this way is shown in figure
A.23. The impulse responses for the operator size 19 × 19 is shown in figure A.24.
Note that the depth image has a perfect circular symmetry but in the middle of the
circle there are some irregular artifacts visible. In the vertical cross-section a ghost
event is observed after and before the main event. An improvement of this method
should aim at an reduction of these artifacts to use this very simple and attractive
218
A.2 2-Dimensional operators for 3-dimensional extrapolation
method to calculate 2-D operators in a more sophisticated way. For example it is
possible to use another (non-linear) interpolation method, or a smooth window to
truncate the operator in the spatial domain.
Error analysis
From an engineering point of view it is interesting to investigate how the different
parameters in the optimization procedure must be chosen to obtain efficient operators which are accurate up to a desired maximum angle. For this analysis it is
necessary to define accuracy in a useful way. In Powell (1981) the most common
used definitions of accuracy are given. In this section we will use the L2 and the
L∞ norms in a certain domain of interest. The domain of interest is defined by
kr < k sin (αmax ) (= krmax ) with k = ωc .
The following L2 and L∞ error norms are defined over the domain of interest
R
ε2 = 
εa∞ =
+
π
4
φ=0
 12
2
kY
(k
)
−
Ŷ
(k
,
φ)k
k
dk
dφ
r
r
r
r
kr =0

R π4 R krmax
2
kY
(k
)k
k
dk
dφ
r
r
r
φ=0 kr =0
R krmax
max
0≤kr ≤krmax
| kY (kr )k − kŶ (kr )k |
max
krmax <kr ≤kN
εp =
"Z
| 1 − kŶ (kr )k |
π
4
φ=0
Z
krmax
kr =0
(A.40)
(A.41)
{if kŶ (kr )k > 1.0}
∂Ep
k
kr k2 dkr dφ
∂kr
# 12
,
(A.42)
with
Ep = arg Y (kr ) − arg Ŷ (kr , φ),
where Ŷ (kr ) is an approximation to the true function Y (kr ), ε2 is the normalized
least-squares error, εa∞ the maximum amplitude error and εp a measurement for the
derivative of the phase error with respect to the polar distance kr . The normalized
ε2 error is a global error and is related to the accuracy of the operator. The amplitude εa∞ error gives an indication of the stability of the operators in a recursive
extrapolation scheme. Note that kY (kr )k2 = 1 in the domain of interest. Included in
the εa∞ error is a stability measurement for k sin (αmax ) < kr ≤ kN . If the amplitude
of the operator in this domain is higher than 1.0 then it contributes to the εa∞ error.
The εp error is defined in such a way that it is sensitive to errors in the circular
symmetry of the operator. In the ideal case the εp error should be zero because of
the circular symmetry of the operator. To compute the derivative with respect to kr
Appendix A: Operator optimization
219
a three point finite difference operator is used to compute the derivative with respect
∂ ∂ky
x
to kx and ky . With these derivatives the ∂k∂ r = ∂k∂x ∂k
∂kr + ∂ky ∂kr is calculated. If
the phase error is large and the derivative with respect to kr is also large then the
operator will have a detectable non-circular character.
To determine the errors due to the recursive use of the operator in a homogeneous
medium (which is a worst case situation) the difference with respect to the reference
impulse response is calculated for every depth slice according to
R
 12
2
xmax R ymax
kY
(x,
y,
z)
−
Ŷ
(x,
y,
z)k
dx
dy
x=0
 .
R y=0
εs (z) = 
(A.43)
xmax R ymax
kY (x, y, z)k2 dx dy
x=0
y=0
This spatial error will be presented in an error curve as function of the angle. The
calculation of the wavenumber errors given by equations (A.40), (A.41) and (A.42)
gives three values for one operator defined for one frequency. For a more useful
definition three frequencies, within the frequency range of interest, are analyzed:
one at a low frequency (in our example 5 Hz), a central frequency (20 Hz) and at
a high frequency (40 Hz). To have a better idea how the different errors in the
operators are exposed in the depth image a number of experiments is carried out
where the most important parameters in the operator design are changing. From
these experiments it is possible to derive an error criterion for the calculated operator
errors which can be used as a measure of accuracy for the extrapolation result.
These experiments are done with different WLSQ optimized operators with a varying
weight function, a change in operator size and a varying maximum angle of interest.
All other parameters remain the same.
The error for three characterizing frequencies is given in table A.3 for different
weighting factors, which are given in the first column. The weighting factor is
defined as the value of the box-shaped weight function outside the domain of
interest. Inside this domain the weight function is given the value 1.0. A practical
Operator
weight
ε2
5 Hz
εa∞
εp
ε2
20 Hz
εa∞
εp
ε2
40 Hz
εa∞
εp
1e-5
5e-5
1e-4
5e-4
1e-3
5e-3
1e-2
5e-2
1.5e-3
2.9e-3
4.1e-3
6.9e-3
7.9e-3
1.0e-2
1.2e-2
2.1e-2
1.9e-2
2.6e-3
3.8e-3
7.4e-3
9.1e-3
1.6e-2
2.0e-2
3.6e-2
4.4e-4
8.4e-4
1.2e-3
1.9e-3
2.0e-3
2.1e-3
2.1e-3
1.6e-3
5.6e-4
1.4e-3
1.7e-3
2.8e-3
3.4e-3
6.0e-3
8.0e-3
1.5e-2
2.8e-2
2.3e-3
3.6e-3
7.3e-3
9.4e-3
1.7e-2
2.4e-2
3.5e-2
8.4e-4
1.6e-3
2.1e-3
2.6e-3
2.7e-3
2.8e-3
3.6e-3
9.4e-4
3.9e-4
8.3e-4
1.3e-3
1.9e-3
2.4e-3
4.0e-3
5.4e-3
1.1e-2
1.3e-2
4.2e-3
5.3e-3
8.1e-3
9.2e-3
1.1e-2
1.3e-2
2.3e-2
1.1e-3
1.4e-3
1.5e-3
2.5e-3
3.2e-3
7.6e-3
1.1e-2
2.2e-2
Table A.3 Error analysis for different weighting factors with a constant operator size (19 × 19) and
maximum angle of interest (60◦ ).
220
A.2 2-Dimensional operators for 3-dimensional extrapolation
-500
-250
x-position [m]
0
250
-500
500
-250
0
250
500
-250
0
250
500
0
-500
250
y-position [m]
-250
0
500
0
250
250
500
z = 220
500
x=0
x=y
a. weightfactor = 1e-3, ε̄2 = 4.2e − 3
-500
-250
x-position [m]
0
250
-500
500
0
-500
250
y-position [m]
-250
0
500
0
250
250
500
z = 220
500
x=0
x=y
b. weightfactor = 5e-2, ε̄2 = 1.7e − 2
Figure A.25 The impulse response as function of the weighting factor. The depth cross section is
equivalent with an angle of 65◦ . Note that for larger weighting factors the error grows
and the result contains more artifacts.
limit of the weighting factor is 5e-5, because smaller factors gives unstable operators
(reflected in the εa∞ error).
• changing weight factor
In figure A.25 two impulse responses are displayed which differ with respect to the
used weighting factor in the operator calculation. The ε̄2 error given in the figures
is the average error over the three characterizing frequencies. From this figure and
Appendix A: Operator optimization
a. amplitude errror
221
b. phase errror
Figure A.26 Amplitude and phase errors for an operator at 40 Hz with a weight factor of 1e-3 and
an operator size of 19 × 19 points. Note the high error values at the edges of the
domain of interest.
table A.3 the following conclusions can de drawn;
(1) For the stable weight factors the ε2 and the εa∞ errors are increasing if the
weighting factor increases. The unstable weight factor (1e-5) is only reflected in the
stability part (second equation on the right hand side of equation (A.41)) of the εa∞
error. The εp and the ε2 error are not sensitive for instabilities outside the domain of
interest. The best weighting factor is therefore that factor which gives an operator
which remains just stable. This factor is easily to determine because it turned out
that this factor remains almost constant for all frequencies of interest with a fixed
operator size.
(2) An average ε2 error smaller than 2e-3 gives an accurate depth image up to the
desired maximum angle. A larger ε̄2 error gives unacceptable artifacts ’inside’ the
main event as observed in figure A.25.
(3) If the ε2 error is small than the other errors are not by definition small too. This
fact can be explained by looking at the amplitude and phase errors which are shown
in figure A.26 for an operator at 40 Hz with a weight factor of 1e-3. The amplitude
error function (a) shows a large peaks at the higher angles, the phase error function
(b) has its largest error at the maximum design angle. In the calculation of the global
ε2 error these peaks are averaged out. In the depth images of figure A.25 these peaks
in the error function are also not visible, but one can imagine that if these peaks
become too big the recursion scheme can become unstable and inaccurate.
• changing operator size
In figure A.27 two different operators are displayed which differ with respect to the
operator size. From this figure, figure A.21 and table A.4 the following conclusions
can de drawn;
222
A.2 2-Dimensional operators for 3-dimensional extrapolation
Operator
size
ε2
5 Hz
εa∞
εp
ε2
20 Hz
εa∞
εp
ε2
40 Hz
εa∞
εp
13x13
19x19
25x25
31x31
37x37
6.1e-3
2.9e-3
1.7e-3
1.3e-3
8.5e-4
3.7e-3
2.6e-3
1.8e-3
1.0e-3
1.1e-3
1.9e-3
8.4e-4
4.4e-4
2.7e-4
1.6e-4
3.1e-3
1.3e-3
6.1e-4
3.5e-4
2.4e-4
3.1e-3
2.3e-3
1.6e-3
1.0e-3
6.8e-4
4.2e-3
1.6e-3
6.3e-4
3.1e-4
2.7e-4
2.7e-3
8.3e-4
4.1e-4
2.5e-4
1.8e-4
4.4e-3
4.2e-3
2.0e-3
7.6e-4
2.3e-2
6.6e-3
1.4e-3
6.8e-4
5.8e-4
4.7e-4
Table A.4 Error analysis for different operator sizes with a constant weighting factor (5e-5) and
maximum angle of interest (60◦ ).
(1) A larger operator size will give more accurate results, but for a certain accuracy
(which is reached for this problem at an operator size of 25 × 25 points with an
average ε2 error in the order of 1e-3) the improvement, by using a larger operator
size, on the result is little. So there exists an optimum efficient operator size.
(2) The large εa∞ error for 40 Hz operator with size 37 × 37 is due to the fact that
the operator has a little amplitude jump after 60◦ on the diagonal kx = ky which is
taken into account by the stability part in the εa∞ error. This effect can be detected
by the small ε2 and εp error. Choosing a slightly bigger maximum angle (65◦ ) will
give a stable operator. The jumps at the edges of the domain of interest are typically
for least squares design methods. These peaks can be suppressed by changing the
weighting function at the edges of the domain or by using an additional optimization
step which uses the results of the first step.
(3) Using a small operator and defining a relative large maximum angle gives
errors which are typically of the form as shown in the vertical cross section of figure
A.27a). These artifacts are due to errors at the higher angles and are most clearly
represented by the ε2 error in table A.4. These artifacts are due to the error peaks
at the edges of the domain of interest. At these critical edges the phase error is
for angles larger than the maximum design angle already large while the amplitude
of the operator is still close to 1. So the large phase error is not suppressed by a
small amplitude error and gives artifacts as observed in the figures. Hoff (1995) has
investigated some methods to suppress these artifacts.
• changing maximum angle
In figure A.28 two different depth images are displayed which differ with respect to
the maximum angle of interest in the operator. The operator size is chosen fixed at
13×13. The result with a maximum angle of 75◦ is unstable and not displayed. From
this figure, figure A.27a and table A.5. The following conclusions can de drawn;
(1) A smaller maximum angle does not automatically gives a better performance for
an operator with the same size. From the εa∞ error in table A.5 it is observed that
there exists an optimum angle which is, for the 13 × 13 operator with the chosen
weight factor (5e-5), an angle between 30 and 45 degrees.
Appendix A: Operator optimization
-500
-250
x-position [m]
0
250
-500
500
223
-250
0
250
500
-250
0
250
500
0
-500
250
y-position [m]
-250
0
500
0
250
250
500
z = 220
500
x=0
x=y
a. operator size = 13x13, ε̄2 = 3.6e − 3
-500
-250
x-position [m]
0
250
-500
500
0
-500
250
y-position [m]
-250
0
500
0
250
250
500
z = 220
500
x=0
x=y
b. operator size = 37x37, ε̄2 = 3.9e − 4
Figure A.27 The impulse response as function of the operator size. The depth cross section is equivalent with an angle of 60◦ . Note that a small operator size gives problems at higher
angles.
(2) The ε2 and εp are of little use because these are not defined for angles outside
the domain of interest.
(3) Note that we kept the weighting factor constant throughout the different experiments, by changing the weight factor it is possible to make a very short operator
which is also stable outside the domain of interest.
From the experiments described above we can develop a criterion which can be
used to determine if a certain operator calculation method gives stable and accurate
224
A.2 2-Dimensional operators for 3-dimensional extrapolation
-500
-250
x-position [m]
0
250
-500
500
-250
0
250
500
-250
0
250
500
0
-500
250
y-position [m]
-250
0
500
0
250
250
500
z = 260
500
x=0
x=y
a. maximum angle = 60◦ ; ε̄2 = 3.7e − 2
-500
-250
x-position [m]
0
250
-500
500
0
-500
250
y-position [m]
-250
0
500
0
250
250
500
z = 360
500
x=0
x=y
◦
b. maximum angle = 45 ; ε̄2 = 3.0e − 3
Figure A.28 The impulse response as function of the maximum angle of interest with an operator
size of 13 × 13. The depth cross section is taken at the maximum angle of interest. Note
that the very small angles give problems.
results in a recursive extrapolation algorithm. To use this criterion the wavenumber
spectrum of the operator must be calculated for three characterizing frequencies and
the ε2 , εa∞ and εp errors have to be calculated for every frequency. These errors must
obey the following relations;
• ε̄2 ≤ 2e−3 accuracy measurement
• εa∞ ≤ 3e−3 stability measurement
• εp ≤ 1e−2 circularity measurement
Appendix A: Operator optimization
225
Operator
angle
ε2
5 Hz
εa∞
εp
ε2
20 Hz
εa∞
εp
ε2
40 Hz
εa∞
εp
15
30
45
60
75
2.1e-3
1.2e-3
2.9e-3
6.1e-3
1.2e-2
3.3e-3
1.5e-3
3.0e-3
3.7e-3
3.1e-3
4.1e-5
4.7e-4
1.9e-3
6.5e-3
7.6e-4
7.2e-4
1.0e-3
3.1e-3
7.8e-3
5.4e-2
1.1e-3
3.2e-3
3.1e-3
3.9e-1
2.0e-5
1.2e-4
3.1e-4
4.2e-3
2.4e-2
6.2e-4
5.3e-4
8.0e-4
2.7e-3
6.14e-3
9.6e-2
6.6e-2
3.0e-3
4.4e-3
1.3e-0
4.9e-5
1.5e-4
4.8e-4
6.6e-3
3.8e-2
Table A.5 Error analysis for different maximum angles with a constant weighting factor (5e-5) and
a fixed operator size of 13 × 13.
The εp error is not tested very well in this section, but in the section about the
McClellan method the εp error turns out to be very useful and the εp criterion can
be defined better.
The performance of the rotated Fourier reconstruction method can be analyzed with
the defined error criteria. In table A.6 the errors are given for different operator sizes.
The 1-D operators are obtained by using the Remez exchange algorithm (similar
results were obtained by using a WLSQ operator). The large ε2 errors indicate that
the overall spectrum of the operator is inaccurate. A better interpolation method
(better than linear) or a smoother window in the spatial domain may improve the
result. With the used linear interpolation method a larger operator size does not
improve the result significantly.
Operator
size
19x19
25x25
31x31
ε2
5 Hz
εa∞
ε
2.4e-2
1.8e-2
1.9e-2
3.9e-2
3.2e-2
3.6e-2
4.0e-3
2.4e-3
2.5e-3
p
ε2
20 Hz
εa∞
ε
5.7e-3
4.8e-3
3.7e-3
1.3e-2
1.3e-2
1.3e-2
6.2e-3
3.8e-3
2.2e-3
p
ε2
40 Hz
εa∞
εp
4.6e-3
2.3e-3
1.6e-3
1.0e-2
1.0e-2
5.1e-3
6.7e-3
1.6e-3
2.8e-3
Table A.6 Fourier reconstructed operators which are accurate up to a maximum angle of 60◦ . The
1-D operators are obtained by using the 1D WLSQ method.
Up to now the errors were calculated in the wavenumber domain and interpreted in
the spatial domain, but with the aid of the reference result and equation (A.43) it
is possible to calculate an error directly in the spatial domain. In figure A.29 this
error is shown for the 19 × 19 and 25 × 25 WLSQ operators which impulse response
is displayed in figure A.21. The vertical cross section is displayed for the diagonal
x = y and the horizontal cross section for an angle of 65◦ . The top picture on the
right hand side shows the error as function of the angle (=depth). From these errors
the following observations are made:
(1) The increasing error line for higher angles as displayed in figure A.29 is due to
226
A.2 2-Dimensional operators for 3-dimensional extrapolation
0.8
-500
-250
x-position [m]
0
250
500
0.6
error
-500
0.2
y-position [m]
-250
0
0◦ 15◦
0
0
250
250
500
z = 220
500
a. 19x19 point operator
-500
-250
x-position [m]
0
30◦
45◦
60◦
65◦
30◦
45◦
60◦
65◦
x=y
0.8
250
500
0.6
error
-500
0.4
0.2
-250
y-position [m]
0.4
0
0◦ 15◦
0
0
250
250
500
z = 220
500
x=y
b. 25x25 point operator
Figure A.29 The spatial error as function of angle (=depth) together with a horizontal (for 65◦ )
and vertical (for x = y) cross section of the error. Note that the all cross sections are
displayed with the same scaling factor.
amplitude errors and artifacts at the higher angles.
(2) Increasing the operator size increases the accuracy; less artifacts and a better
amplitude.
(3) From the impulse response alone it is difficult to interpret the accuracy of the
operator, comparing it with the reference operator gives a good indication of the
errors in the operator and the influence of the recursive application of the operators.
Given the error criteria we can also determine how the weight factor and the operator
Appendix A: Operator optimization
227
size must be chosen for a maximum angle of interest. The results of these experiments
are summarized in table A.7. The small angles 15◦ and 30◦ are difficult to optimize
for the given maximum angle, but by choosing a slightly bigger angle the operator
can become stable and accurate for the smallest operator size possible. For example
to get the operator for 30◦ a maximum design angle of 40◦ degrees has to be chosen.
For the higher angles this problem does not occur. If one wants to design operators
with small maximum angles of interest and suppression of all the higher angles a
larger operator size must be chosen than the one given in table A.7. The larger
ε̄p error in the 75◦ operators is not as bad as it looks, the largest error peaks are
positioned at the edges of the domain of interest.
Note that for small operator sizes the higher frequencies are most sensitive to errors,
for the larger operators the lower frequencies are more sensitive to errors. This
behavior is related to the WLSQ optimization method. A very small operator has a
limited number of ’error’ peaks in the frequency domain due to the limited number
of contributing wavenumber components. The WLSQ optimization method with a
limited number of wavenumber components cannot have very large peaks (Berkhout,
1984). If there are more wavenumber components the WLSQ method can build up
large peaks in the error function (Gibbs phenomenon).
angle
15
30
45
60
75
Operator
size
weight
5x5
9x9
13x13
19x19
31x31
1e-5
2e-5
4e-5
4e-5
6e-5
ε2
20 Hz
εa∞
εp
ε̄2
Average
ε̄a∞
ε̄p
1.6e-3
1.4e-3
1.0e-3
1.2e-3
1.7e-3
3.2e-3
3.0e-3
3.3e-3
2.4e-3
1.2e-3
1.6e-5
5.6e-4
3.5e-4
1.6e-3
4.8e-3
2.4e-3
1.9e-3
1.5e-3
1.5e-3
1.9e-3
2.2e-3
2.9e-3
3.1e-3
2.9e-3
1.4e-3
7.1e-5
2.7e-4
3.9e-4
1.2e-3
4.3e-3
Table A.7 Optimum operators which are accurate up to a maximum angle of interest. Note that for
small angles these operators are stable but don’t suppress all higher angles.
228
A.2.2
A.2 2-Dimensional operators for 3-dimensional extrapolation
McClellan transformation, expansion in cos (kr )
The McClellan transformation transforms a 1-D convolution operator to a 2-D convolution operator with a certain symmetry. This transformation is of interest because the implementation is simple and the computation of the transformation coefficients can be done efficiently. Hale (1991a) introduced the McClellan transformation
into the Geophysical world and described two related techniques which replace the
2-dimensional direct spatial convolution: (1) transformation of the non-recursive
1-dimensional symmetrical filter in a 1-dimensional recursive filter by using the
Chebychev recursion formula, (2) the McClellan transformation of a 1-dimensional
filter to a circular symmetric 2-dimensional filter. In this subsection first the transformation from a 1-dimensional filter to a 2-dimensional filter will be discussed.
Next the Chebychev recursion formula is explained and at the end of this section
several methods are discussed which optimize the steps and coefficients used in the
McClellan transformation.
McClellan Transformation form 1-D to 2-D
If the operator has a circular symmetry it is possible to reduce the computation
time of the 2-dimensional filter by means of a McClellan transform. The McClellan
transform (McClellan, 1973) defines a mapping from a 1-D wavenumber axis to the
2-D wavenumber domain. The change of variables to be described depends on the
fact that both the operator approximations in 1- and 2-dimensions can be written as
sums of cosine functions. The 1-dimensional filter problem of an even symmetrical
operator can be rewritten as (note the similarity with equation (A.31))
Ỹ (kr ) ≈ Y (0) + 2
Ỹ (kr ) ≈
Ỹ (kr ) ≈
M
X
M
X
Y (m∆x) cos (kr m∆x),
(A.44)
m=1
Y ′ (m∆x) cos (kr m∆x),
(A.45)
Ŷ (m∆x) cos (kr ∆x)m ,
(A.46)
m=0
M
X
m=0
with the choice of a suitable set of coefficients Ŷ (m∆x) that approximate the 1-D
extrapolation operator Ỹ (kr ). In equation (A.45) Ym′ = 2Ym for m = 1, . . . M and
Ym′ = Ym for m = 0. The step from equation (A.45) to equation (A.46) can be seen
by letting φ = cos (kr ∆x), then cos (kr m∆x) = cos (m arccos (φ)) = Tm (φ), where
Tm (φ) is the Chebychev polynomial of order m as introduced in equation (A.21).
Each cosine term in equation (A.45) may then be expressed in the form
cos (kr m∆x) =
N
X
n=0
αm,n cos (kr ∆x)m ,
(A.47)
Appendix A: Operator optimization
229
Figure A.30 The wavenumber spectrum of a first order McClellan operator with a 1-D operator of
19 spatial point, 128 × 128 kx , ky points with c = 1000 m/s, f = 25 Hz, ∆x =
∆y = ∆z = 10 m and a maximum angle of interest of 65◦ .
where the αm,n are real coefficients and easily obtained with the Chebychev recursion
formula. Equation (A.45) reduces then further to
Ỹ (kr ) ≈
M
X
Y ′ (m∆x)Tm (φ) =
M
X
Ŷ (m∆x)φm ,
(A.48)
m=0
m=0
where both right-sides of the equation are now polynomials in φ.
The cosine terms in equation (A.46) can be approximated by a 2-dimensional filter
(assuming ∆x = ∆y)
cos (kr ∆x) ≈
Q
P X
X
cpq cos (kx p∆x) cos (ky q∆y),
(A.49)
p=0 q=0
where cpq are called the McClellan factors after McClellan (1973). By making the
substitution of equation (A.49) into equation (A.46) it reduces to
Ỹ (kr ) ≈
M X
N
X
Ÿ (m∆x, n∆y) cos (kx ∆x)m cos (ky ∆y)n ,
(A.50)
m=0 n=0
which can be put in the form (using Chebychev’s recursion formula again)
Ỹ (kr ) ≈
M X
N
X
Y̆ (m∆x, n∆y) cos (kx m∆x) cos (ky n∆y),
(A.51)
m=0 n=0
which is the desired form for a 2-dimensional filter which was already shown in
equation (A.31) (Note that Ÿ and Y̆ are scaled versions of Y ). For example for
230
0
0.1
k_y cycles
0.2
0.3
0.4
0.5
0
0.1
0.1
0.2
0.2
k_x cycles
k_x cycles
0
A.2 2-Dimensional operators for 3-dimensional extrapolation
0.3
0.4
0.5
0
0.1
k_y cycles
0.2
0.3
0.4
0.5
0.3
0.4
first order McClellan
0.5
second order McClellan
Figure A.31 Contour plots for the first and second order McClellan transformation. Note that for
higher wavenumbers both approximations deviate from the ideal circular (dashed) line.
P = Q = 1 the transformation for circular symmetry reduces to a 9-term McClellan
convolution operator (also called a first order approximation) which is given by Hale
(1991a) where −c00 = c10 = c01 = c11 = 0.5 and
cos (kr ∆x) ≈ −1 + 0.5(1 + cos (kx ∆x))(1 + cos (ky ∆y)).
(A.52)
In figure A.30 an extrapolation operator is shown which is designed with the first
order McClellan transformation of equation (A.52) and a 1-D operator Y (m∆x) of
10 points (19 point full operator length). Note the square shape of the operator for
the wavenumbers near the Nyquist wavenumber (the edges of the figure). Choosing
P = Q = 2 gives a 17-term McClellan transform (second order approximation),
which is also given by Hale (1991a)
cos (kr ∆x) ≈ −1 + 0.5(1 + cos (kx ∆x))(1 + cos (ky ∆y))
− 0.5c(1 − cos (2kx ∆x))(1 − cos (2ky ∆y)),
(A.53)
with c = 0.0255. The McClellan factors cpq can be derived by defining points in
the kx , ky plane which map to a point on the k-axes of the 1-D operator such that
all coefficients are uniquely defined (for the first order McClellan transform only 4
mapping points are needed). The problem with the McClellan transform, given the
original McClellan factors in equations (A.52) and (A.53), is that for higher angles
the transformation deviates from the ideal circular shape. The contour plots shown
in figure A.31 represent the contours of the first order McClellan (P = Q = 1)
and the second order McClellan (P = Q = 2) transformation. In the contour plots
Appendix A: Operator optimization
231
the deviation from the ideal circular shape for the higher wavenumbers is observed
clearly. The second order transformation reduces the deviation a little but remains
still significant.
Chebychev recursion formula
The second improvement in the computation scheme is the transformation of the
non-recursive 1-dimensional filter to a 1-dimensional recursive filter derived from the
recursive formula of the Chebychev polynomials
cos (mφ) = 2 cos (φ) cos ((m − 1)φ) − cos ((m − 2)φ).
(A.54)
This Chebychev filter structure is not useful for 1-dimensional filters. Direct convolution is both simpler and more efficient. The Chebychev structure is more advantageous for 2-dimensional operators with an even symmetry, such as the circular
symmetric extrapolation operators. Writing equation (A.54) for the first four terms
in the 1-dimensional case gives
PM
(A.55)
Ỹ (kx ) = Y0 + 2 m=1 Ym cos (kx m)
= Y0
+2Y1 [cos(kx )]
+2Y2 [2 cos(kx ) cos(kx ) − 1]
+2Y3 [2 cos(kx ) (2 cos(kx ) cos(kx ) − 1) − cos(kx )]
+2Y4 [2 cos(kx ) 2 cos(kx ) (2 cos(kx ) cos(kx ) − 1) − cos(kx ) − (2 cos(kx ) cos(kx ) − 1)]
and is shown in figure A.32. This recursive scheme can be implemented in the
computer without much effort. McClellan and Chan (1977) have analyzed this so
called Chebychev structure in detail and observed that the scheme requires the
minimum number of multiplications in comparison with the direct scheme and it
is the most stable scheme with respect to the round-off noise. Note that only the
coefficients of the 1-D operator are involved. Hence, the number of computations
−
Pi
h(x,y)
2h(x,y)
−
2h(x,y)
2h(x,y)
−
F0
2F1
+
2F2
+
2FM
+
Pi+1
Figure A.32 Chebychev recursion scheme. The h(x, y) boxes represent the 2-D McClellan transformation of cos (kr ), Ym represents the coefficients of the 1-D convolution operator.
232
A.2 2-Dimensional operators for 3-dimensional extrapolation
-500
-250
x-position [m]
0
250
-500
500
-250
0
250
500
-250
0
250
500
0
-500
250
y-position [m]
-250
0
500
0
250
250
500
z = 220
500
x=0
x=y
a. 13x9 McClellan operator
-500
-250
x-position [m]
0
250
-500
500
0
-500
250
y-position [m]
-250
0
500
0
250
250
500
z = 220
500
x=0
x=y
b. 13x17 McClellan operator
Figure A.33 Depth images obtained with the original McClellan transformation. For the higher
angles both operators deviate from the ideal circular shape. Note the deviation at the
higher angles in the diagonal slice.
depends linearly on the length N of the 1-D operator and not on N 2 as in the
implementation of a direct 2-D convolution. The computation times for several 1D operator lengths and different McClellan operators are given at the end of this
section.
In figure A.33 the migration results of the McClellan transformation combined with
the Chebychev recursion scheme are shown for both the first (a) and second (b)
order McClellan transformation. The used optimized 1-D convolution operators
Appendix A: Operator optimization
233
have a full length of 25 points and are obtained with the Remez exchange algorithm.
The Chebychev recursion scheme makes use of the even symmetry in the operator
so only 13 points of the 1-D operator are needed as expansion terms in the recursive
scheme. The notation of the operator size used in figure A.33 gives first the number
of terms in the expansion and second the size of one single term. For example the
notation 13 × 9 means that the expansion is done in 13 terms where every term
consists of an operator with 9 points.
The cross-sections in figure A.33 give a good illustration how the McClellan transformation handles the higher angles. Note that the deviation of the ideal circle for
the second order (17 term) is only a little less than for the first order (9 term) McClellan transformation. The noise around the source position in the vertical slices are
an artifact of the used 1-D operator. Despite the deviation at the higher angles the
McClellan transformation combined with the Chebychev recursion scheme is a very
powerful and useful approach. The performance at higher angles can be improved
in several ways. In the the following paragraphs four improvements are discussed.
Hazra and Reddy Coefficients
Optimizing the design of the McClellan factors cpq in equation (A.49) is a good
way to improve the performance of the McClellan transformation. The first order
filter is of special interest because it is a small and therefore fast operator. The
aim of the technique proposed by Hazra and Reddy (1986) is to make the maximum
contour of interest of the 2-D operator approximate a circle with a high degree of
accuracy. This better approximation is achieved by mapping an additional point,
defined by the maximum wavenumber of interest krmax , of the 2-D operator onto
the maximum wavenumber of interest (kxmax ) of the 1-D filter. This mapping of the
additional point is obtained by making the wavenumber kxmax of the 1-D filter as one
of the design parameters. A consequence of this is that the maximum wavenumber
of interest of the 2-D operator on the kx -axis and ky -axis may be different from the
maximum wavenumber of interest of the 1-D filter.
The first order original McClellan transformation maps the origin (0, 0) in the
(kx , ky ) plane onto the point kx = 0 of the kx -axis in the wavenumber response of
the 1-D operator. The points (kx,N , 0), (0, ky,N ) and (kx,N , ky,N ) from the (kx , ky )
plane all map onto kx = kx,N . With the definition of these four points the coefficients in the first order McClellan transform are uniquely determined. This mapping
has the following properties: (1) the contours of the McClellan transformation are
approximately circular for low values of kr and deviates considerably from circular
contours as kr increases and is square at kr = kN yquist , (2) the original McClellan
transformation makes the frequency response of the 2-D operator along the kx -axis
and along the ky -axis identical to the frequency response of the original 1-D operator. The contour plots in figure A.31 show that the deviation from the circular
contour is maximum near the neighborhood of the diagonal joining the points (0, 0)
234
0
k_y cycles
0.2
0.3
0.1
0.4
0.5
0
0.1
0.1
0.2
0.2
k_x cycles
k_x cycles
0
A.2 2-Dimensional operators for 3-dimensional extrapolation
0.3
0.4
0.5
0
0.1
k_y cycles
0.2
0.3
0.4
0.5
0.3
0.4
optimized for 20 Hz (= 0.2 cycles)
0.5
optimized for 40 Hz (= 0.4 cycles)
Figure A.34 Contour plots of the optimized Hazra and Reddy transformation for two different frequencies. Note the improvement in the circular shape in comparison with the original
McClellan transformation.
and (kx,N , ky,N ) in the (kx , ky ) plane. It is possible, for a given maximum krmax ,
to improve the contour by forcing an appropriate point on this diagonal to be on
the circular contour. With this mapping of an extra point on the circular contour,
it is not possible, to make the frequency response of the 1-D operator identical to
the frequency response of the original 1-D operator along the kx -axis and along the
ky -axis. Thus the mapping of the extra point on the diagonal is only possible when
the maximum wavenumber of interest along the kx axes is one of the design parameters. The McClellan factors according to Hazra and Reddy are dependent on the
maximum wavenumber of interest and are given by
c11 =
ab − 2ac
1
, c01 = c10 = − c11 , c00 = 1 − c01 − c10 − c11 ,
2
2bc
g
where
b − 2c
b
k max
krmax
, a=
), c = sin2 ( r√ ), g = 2 +
2
2
c
g
2 2
−1 √
= 2 sin ( a)
b = sin2 (
kxmax
(A.56)
where krmax is the maximum wavenumber of interest of the circular symmetric 2-D
filter and kxmax the maximum wavenumber of interest of the 1-D operator. For a
more detailed discussion on the derivation of the parameters in equation (A.56) the
reader is referred to Hazra and Reddy (1986). In figure A.34 two contour plots
are shown for two different frequencies. Up to the desired maximum value these
contours are circular, outside the desired value the contours are not circular shaped.
Appendix A: Operator optimization
235
Figure A.35 The wavenumber spectrum of a first order McClellan operator with the optimized
Hazra and Reddy coefficients and a 1-D operator of 19 spatial points with 128 × 128
kx , ky points with c = 1000 m/s, f = 25 Hz, ∆x = ∆y = ∆z = 10 m and a
maximum angle of interest of 65◦ .
The operator calculated with these optimized contours is shown in figure A.35. Note
the difference with figure A.30.
In the extrapolation algorithm first the McClellan factors are calculated according
to A.56 for a given krmax which gives besides the optimized McClellan factors also
a kxmax for the 1-D operator. With this calculated kxmax value the 1-D operator is
designed. Note that kxmax is always smaller than krmax . Due to the choice of the
coefficients this 1-D operator is stretched to the desired 2-D operator. To compensate
for this stretch the 1-D convolution operator must be calculated with a scaled ∆z.
This can be explained by regarding the effect of the optimized Hazra and Reddy
factors as a scaling to the kx -axis. The 1-D phase shift operator is then given by
1
Ỹ (kx ) = exp (−j [k 2 − (αkx )2 ] 2 ∆z)
1
k
= exp (−j [( )2 − kx2 ] 2 α∆z)
α
where
k
α
(A.57)
= kxmax and α∆z is the scaled depth step.
The migration results for the Hazra and Reddy optimized coefficients are shown in
figure A.36. The circular shape of the depth slice is good and the artifacts at the
diagonals, which were present in the original McClellan transformation, are absent.
Note that the computation time of the optimized coefficients is very small, only a
few multiplications and additions per frequency, so the same computation effort is
required as for the original McClellan transformation which makes the method very
attractive. It will be interesting to investigate if it is possible to adjust the 17-points
McClellan transformation with the same method.
236
A.2 2-Dimensional operators for 3-dimensional extrapolation
-500
-250
x-position [m]
0
250
-500
500
-250
0
250
500
0
-500
250
y-position [m]
-250
0
500
0
250
250
500
z = 220
500
x=0
x=y
13x9 Hazra and Reddy operator
Figure A.36 Depth images obtained with the optimized Hazra and Reddy transformation. Note the
improvement in the circular shape in the depth slice in comparison with the original
McClellan transformation.
Optimized McClellan factors
The aim in optimizing the McClellan transformation is to choose the McClellan
factors cpq in equation (A.49) such that the contours produced by the transformation
have some desired shape. For some examples it is sufficient to control the shape of
one single contour. In other problems it is necessary to design the shape of the
contours in a specific part of the wavenumber domain (Mersereau et al., 1976). The
error function which has to be optimized for a circular contour design is given by
E = M (cpq , kx , ky ) − cos
M (cpq , kx , ky ) =
Q
P X
X
q
kx2 + ky2 with
cpq cos (kx ) cos (ky )
(A.58)
p=0 q=0
Equation (A.58) is a non-linear function of the unknown parameters, so a computation intensive non-linear optimization scheme must be used for the minimization.
However, it is possible to reformulate the problem as a linear approximation problem
to arrive at a sub optimum solution, Mersereau et al. (1976). In the example shown
in this section a non-linear optimization scheme (CFSQP, Tits and Zhou (1996)) is
used. With this scheme several contours are optimized within the band of interest
Appendix A: Operator optimization
237
Figure A.37 The wavenumber spectrum of a second order McClellan operator with optimized coefficients and a 1-D operator of 19 spatial points with 128 × 128 kx , ky points with
c = 1000 m/s, f = 20 Hz, ∆x = ∆y = ∆z = 10 m and a maximum angle of
interest of 65◦ .
and the constraint
|
Q
P X
X
p=0 q=0
cpq cos (kx p∆x) cos (ky q∆y)| ≤ 1
0 ≤ kx ≤ kx,N (=
(A.59)
π
π
), 0 ≤ ky ≤ ky,N (=
)
∆x
∆y
is put for all points of the mapping in the (kx , ky ) plane. The contours to be
optimized in the objective function of equation (A.58) are defined by the maximum
wavenumber value of interest. With this definition of the optimization problem the
first order McClellan transformation cannot be optimized any further, but the second
order transformation (with an expansion to 25 points, which means that all cross
terms within the second order are used) can be improved. In the implementation
for the optimization of the McClellan operators explicitly use is made of the circular
symmetry in the McClellan operator if ∆x = ∆y. In figure A.38 the migration pulse
response is shown for the optimized 25-point McClellan operator, with frequency
dependent coefficients and a 1-D operator of 25 points (13 terms). The non-linear
computation time can be reduced by calculating the optimized coefficients for a
wavenumber range instead of every wavenumber. In the shown example only four
sub domains of the total wavenumber domain of interest are used to keeps the time
to compute the McClellan factors small. It also possible to optimize the coefficients
independent of the wavenumber frequency or for different shaped McClellan filters,
see Blacquière (1991).
238
A.2 2-Dimensional operators for 3-dimensional extrapolation
-500
-250
x-position [m]
0
250
-500
500
-250
0
250
500
0
-500
250
y-position [m]
-250
0
500
0
250
250
500
z = 220
500
x=0
x=y
13x25 optimized McClellan operator
Figure A.38 Depth slice at z = 220 m, a vertical slice for x=0 (top right) and bottom right a vertical
slice for x=y with optimized McClellan coefficients. Note the circular shape and the
small artifacts.
Rotated Coefficients
Biondi and Palacharla (1994) describe a method which reduces the error of the deviation of the circle at the diagonal in the (kx , ky ) plane by using a rotated McClellan
operator, the rotation angle being π4 . In the downward extrapolation scheme the
rotated McClellan operator and the original McClellan filters are alternately used
as convolution operator. The convolution with the rotated McClellan operator can
be implemented in an efficient way. For a more detailed discussion the reader is
referred to Biondi and Palacharla (1994).
Series expansion in cos (kr ∆x)
All improvements described thus far make use of the Chebychev recursion scheme,
but it is also possible to use a direct expansion in cos (kr ∆x). To see the difference
between the two schemes the Chebychev recursion scheme and the direct scheme are
given below
Ỹ (kr ) ≈
≈
≈
M
X
Ym Tm (cos (kr ∆x))
(A.60)
am cos (kr ∆x)m
(A.61)
âm h̃m (kx , ky )
(A.62)
m=0
M
X
m=0
M
X
m=0
Appendix A: Operator optimization
Pi
h(x,y)
â0
h(x,y)
â1
+
239
h(x,y)
â2
+
âM
1
+
âM
+
Pi+1
Figure A.39 Direct scheme for series expansion in cos (kr ∆x). Note the simple structure in comparison with the Chebychev recursion scheme.
with
h̃(kx , ky ) ≈ cos (kr ∆x)
(A.63)
In equation (A.62) h̃ (defined in equation (A.63)) is optimized first and with this
approximation to cos (kr ∆x) âm is optimized. In equation (A.60), the Chebychev recursion scheme, the optimization of cos (kr ∆x) is independent of the optimization of
Ym . The series expansion of equation (A.62) will in principle give a better operator,
because the optimization of the series coefficients is dependent on the approximation
to cos (kr ∆x). To optimize the expansion factors âm the WLSQ method is used.
To approximate cos (kr ∆x) the McClellan or Hazra and Reddy transformation can
be used, but it is also possible to use the WLSQ optimization method as described
before.
The recursive convolution scheme of equation (A.62) is given in figure A.39. This
scheme is less complicated to implement in the computer and more important it can
be optimized better by the compiler and user. In figure A.40a the impulse response
is shown with the McClellan coefficients of equation (A.52) for the approximation to
cos (kr ∆x). The number of expansion terms is chosen equal to 13. In figure A.40b
the series expansion in cos (kr ∆x) is done with an approximation to cos (kr ∆x)
obtained with the WLSQ method introduced in section A.2.1. Note that with a 9
point approximation to cos (kr ∆x) the result is accurate up to the higher angles.
In the next subsection a detailed error analysis is given for all discussed McClellan
methods.
Error analysis
Using the analysis technique, which was introduced by the direct method, it is possible to analyze the performance of the different McClellan transformations. In table
A.8 the errors are given for five types of McClellan transformations; the original McClellan transformation in first (*x9) and second (*x17) order, the optimized Hazra
& Reddy factors (HR) in first order, the non-linear optimized factors in the expanded second order (*x25) and the series expansion in cos (kr ∆x) with the Hazra and
Reddy coefficients and the WLSQ operators. From the results in the table and fig-
240
A.2 2-Dimensional operators for 3-dimensional extrapolation
-500
-250
x-position [m]
0
250
-500
500
-250
0
250
500
-250
0
250
500
0
-500
250
y-position [m]
-250
0
500
0
250
250
500
z = 220
500
x=0
x=y
a. 13x9 with original McClellan coefficients
-500
-250
x-position [m]
0
250
-500
500
0
-500
250
y-position [m]
-250
0
500
0
250
250
500
z = 220
500
x=0
x=y
b. 13x9 with WLSQ optimized coefficients
Figure A.40 Impulse response of the expansion in cos (kr ∆x) with original McClellan and WLSQ
optimized coefficients for cos (kr ∆x). The series coefficients in all examples are optimized by using the WLSQ method.
ures A.33, A.36, A.38, A.40 the following remarks can be made;
(1) For higher frequencies the εp error in the original McClellan transformation increases significantly for both the first and second order approximation.
(2) The differences in pulse responses between the methods with, the original McClellan factors, the Hazra & Reddy factors, the non-linear optimized factors or the
series expansion can be determined from the εp error. In figure A.41 the phase error
and the ∂k∂ r of the phase error, as used in equation A.42, for the different methods
Appendix A: Operator optimization
241
Operator
McClellan
ε2
5 Hz
εa∞
εp
ε2
20 Hz
εa∞
εp
ε2
40 Hz
εa∞
εp
10x9
13x9
16x9
10x17
13x17
16x17
1.7e-3
1.4e-3
1.1e-3
1.7e-3
1.3e-3
1.1e-3
1.0e-3
1.2e-3
1.3e-3
1.0e-3
1.2e-3
1.3e-3
4.5e-4
3.2e-4
2.3e-4
4.4e-4
3.1e-4
2.3e-4
6.6e-3
6.3e-3
6.3e-3
1.8e-3
1.1e-3
7.8e-4
2.2e-3
1.7e-3
1.1e-3
2.4e-3
2.0e-3
1.4e-3
4.8e-3
3.9e-3
3.7e-3
2.7e-3
1.4e-3
9.7e-4
5.1e-2
5.2e-2
5.1e-2
2.4e-2
2.4e-2
2.4e-2
1.5e-3
1.4e-3
1.6e-3
1.5e-3
1.6e-3
1.9e-3
3.8e-2
3.8e-2
3.9e-2
2.2e-2
2.3e-2
2.3e-2
HR
10x9
13x9
16x9
2.6e-3
1.2e-3
1.4e-3
5 Hz
1.2e-3
6.7e-4
1.4e-3
8.1e-4
2.9e-4
3.2e-4
7.2e-3
7.1e-3
7.2e-3
20 Hz
1.7e-3
2.7e-3
1.7e-3
2.8e-3
2.3e-3
1.9e-3
8.2e-2
8.2e-2
8.2e-2
40 Hz
7.7e-3
1.6e-2
2.7e-2
1.8e-2
1.8e-2
1.8e-2
Optimized
10x25
13x25
16x25
1.7e-3
1.4e-3
1.1e-3
5 Hz
1.0e-3
1.2e-3
1.3e-3
4.4e-4
3.2e-4
2.3e-4
1.8e-3
1.2e-3
9.2e-4
20 Hz
2.4e-3
2.0e-3
1.4e-3
2.6e-3
1.4e-3
1.0e-3
7.2e-3
7.2e-3
7.2e-3
40 Hz
1.5e-3
1.9e-3
2.3e-3
9.4e-3
8.6e-3
8.5e-3
Series HR
10x9
13x9
16x9
5.2e-3
2.0e-3
1.5e-3
5 Hz
2.5e-3
1.5e-3
7.7e-4
1.6e-3
5.3e-4
3.6e-4
1.7e-3
9.8e-4
6.4e-4
20 Hz
1.8e-3
1.3e-3
1.2e-3
2.5e-3
1.4e-3
8.4e-4
8.4e-3
8.4e-3
8.4e-3
40 Hz
2.3e-3
2.4e-3
2.2e-3
6.7e-3
5.6e-3
5.2e-3
Series WLSQ
10x9
13x9
16x9
10x25
13x25
16x25
2.8e-3
1.3e-3
5.3e-4
2.2e-3
1.3e-3
5.7e-4
5 Hz
1.8e-3
6.2e-4
3.4e-4
1.2e-3
4.1e-4
2.7e-4
8.2e-4
3.4e-4
1.1e-4
6.4e-4
3.2e-4
1.2e-4
5.1e-3
5.0e-3
5.0e-4
8.3e-4
4.3e-4
1.9e-4
20 Hz
1.7e-3
1.4e-3
4.1e-4
8.4e-4
6.9e-4
2.9e-4
3.8e-3
3.4e-3
3.2e-3
1.2e-3
5.9e-4
2.6e-4
2.2e-3
2.0e-3
1.9e-3
1.4e-3
1.2e-3
1.1e-3
40 Hz
2.3e-3
1.6e-3
4.2e-4
1.6e-3
1.2e-3
3.4e-3
4.5e-3
3.6e-3
3.4e-3
2.8e-3
1.5e-3
1.4e-3
Table A.8 Errors in the extrapolation operators for; the original McClellan transformation in first
(*x9) and second (*x17) order, the optimized Hazra & Reddy factors (HR) in first order,
non-linear optimized factors in the expanded second order and the series expansion in
cos (kr ) with the Hazra and Reddy coefficients and WLSQ operators. The maximum
angle of interest is 60◦ .
are displayed for a frequency of 40 Hz, a maximum angle of 60◦ and 13 terms in the
expansion. The second order method with the original McClellan factors give a rapidly increasing phase error where the largest errors occurs at the diagonal from (0, 0)
to (kx,N , ky,N ). The phase error of the Hazra & Reddy method is less rapidly increasing and the smallest error is positioned at the diagonal from (0, 0) to (kx,N , ky,N ).
The non-linear optimized method gives error peaks at the edges of the domain of
242
A.2 2-Dimensional operators for 3-dimensional extrapolation
a. Phase error in original McClellan 13x17
Radial derivative of (a)
b. Phase error in Hazra & Reddy 13x9
Radial derivative of (b)
c. Phase error in Non-linear optimization 13x25
Radial derivative of (c)
d. Phase error in Series expansion 13x9
Radial derivative of (d)
Figure A.41 Phase errors and the radial derivative of the phase error for the different McClellan
methods. Note the different scales on the vertical axes.
Appendix A: Operator optimization
243
0.8
-500
-250
x-position [m]
0
250
500
0.6
error
-500
0.2
y-position [m]
-250
0
0◦ 15◦
0
0
250
250
500
z = 220
500
a. 13x17 original McClellan
-500
-250
x-position [m]
0
250
30◦
45◦
60◦
65◦
30◦
45◦
60◦
65◦
x=y
0.8
500
0.6
error
-500
0.4
0.2
-250
y-position [m]
0.4
0
0◦ 15◦
0
0
250
250
500
z = 220
500
x=y
b. 13x9 series in cos (kr )
Figure A.42 Error in the impulse response of the Chebychev expansion in cos (kr ∆x) with original McClellan (a) and for a series expansion in cos (kr ∆x) with WLSQ optimized
coefficients (b).
interest and has the largest error at the diagonal from (0, 0) to (kx,N , ky,N ). The
series expansion method with a WLSQ approximation to cos (kr ∆x) gives the smallest errors with error peaks at the edges on kx and ky axes.
(4) The approximation to cos (kr ∆x) can be done with many different methods.
Crucial in the performance of the operator is that the coefficients in the expansion
(Chebychev or series) are optimized by using the approximation to cos (kr ∆x).
In figure A.42 the spatial error is given for a horizontal cross section at 65◦ and a
244
A.2 2-Dimensional operators for 3-dimensional extrapolation
vertical cross section at x = y for both the original second order McClellan transformation (13x17) and the series expansion in cos (kr ∆x) (13x9). The original McClellan scheme gives dispersive artifacts for the higher angles. The expansion in
cos (kr ∆x) (with less coefficients in the approximation to cos (kr ∆x)) does not have
these artifacts but is less accurate in amplitude for the higher angles.
In table A.9 the shortest accurate operator is given as function of the maximum
angle of interest. For small angles the original first order McClellan transformation
in Chebychev series gives already good results. For intermediate angles the second
order McClellan scheme or the first order Hazra & Reddy factors are sufficient, for
higher angles the series expansion with WLSQ optimized series coefficients gives the
best results. The method of Hazra & Reddy cannot be found back in table A.9
because of the large ε2 error at the higher frequencies caused by the stretching of
the operator. For the 75◦ angle a 5x5 approximation to cos (kr ∆x) is needed.
angle
15
30
45
60
75
Operator
size
method
4x9
5x9
7x17
10x9
18x25
McC
McC
McC
Series
Series
ε2
40 Hz
εa∞
εp
ε̄2
Average
ε̄a∞
ε̄p
1.7e-3
3.9e-3
5.2e-3
2.2e-3
6.9e-3
2.9e-3
2.0e-3
2.5e-3
2.3e-3
2.6e-3
1.5e-4
1.0e-3
3.3e-3
4.5e-3
2.6e-2
1.8e-3
3.1e-3
3.1e-3
3.4e-3
4.7e-3
2.0e-3
2.0e-3
2.0e-3
1.9e-3
1.9e-3
1.0e-4
5.4e-4
1.5e-3
3.0e-3
1.2e-2
Table A.9 Optimum operators which are accurate up to a maximum angle of interest.
A.2.3
Expansion in kz
In the previous subsections the direct convolution and the McClellan transformation were explained and impulse responses were shown for different operators. The
McClellan transformation uses the coefficients of a 1-D convolution operator and
approximates the 2-D Fourier components with an optimum filter. However, it is
also possible to approximate the phase shift operator with an expansion other than
the cosine terms of the Fourier transformation. Writing space-frequency wavefield
extrapolation in an operator notation Berkhout (1982).
P + (zm+1 ) = W + (zm+1 , zm ) ∗ P + (zm )
(A.64)
where W + (zm+1 , zm ) is the propagation operator and P + (zm ) is the downgoing
wavefield at depth level zm . In this notation the most simple approximation to
finite difference wavefield extrapolation is made by a single Taylor series expansion.
Appendix A: Operator optimization
245
In the spatial domain with ∆z = zm+1 − zm this approximation is given by
P + (zm+1 ) ≈ P + (zm ) +
∆z ∂P + (zm ) ∆z 2 ∂ 2 P + (zm ) ∆z 3 ∂ 3 P + (zm )
+
+
+ ...
2
3
1!
∂zm
2!
∂zm
3!
∂zm
(A.65)
The extrapolation scheme given in (A.65) can be divided into two parts; one part
i
deals with the estimation of the derivatives ∂z∂ i with respect to zm and the other
m
part deals with the prediction with the aid of the Taylor series. In the wavenumber
+
domain this approximation, with ∂∂zP̃m = −jkz P̃ + , is defined as
jkz ∆z +
(kz ∆z)2 +
(kz ∆z)3 +
P̃ (zm ) −
P̃ (zm ) + j
P̃ (zm ) + . . .
1!
2!
3!
(A.66)
This truncation of the series expansion is an approximation of the phase shift operator by a polynomial in kz , according to
P̃ + (zm+1 ) ≈ P̃ + (zm ) −
2
exp (−jkz ∆z) ≈ 1 − j ∆zkz +
(j ∆z)
3
(kz )2 + O(kz )
2!
(A.67)
The coefficients in the series expansion can be obtained by using the constants in
the Taylor series as given in equation (A.67) or by using an optimization technique.
In the next two subsections two different optimization methods are used; the WLSQ
method with the L2 norm and the Remez exchange method with the L∞ norm.
Series expansion in kz with L2 -norm
The advantage of an expansion in kz is that if the kz operator can be approximated
by a short spatial convolution operator, and the number of terms in the series expansion of the phase shift operator remains small, the computation time can be reduced
in comparison with the direct 2-D convolution. To arrive at the direct spatial convolution scheme, given in figure A.43, the kz operator is transformed in an optimum
way to the space domain and applied several times to the data. Every time a 2-D
convolution (indicated by the box H1 ) with the spatial kz operator is carried out on
Pi
H1
â0
H1
â1
+
H1
â2
+
âM
1
+
âM
+
Pi+1
Figure A.43 Series expansion in terms of kz . The H1 boxes represent the 2-D convolution with the
optimized spatial kz ≈ H̃1 operator, the âm represent the optimized coefficients in the
series expansion.
246
A.2 2-Dimensional operators for 3-dimensional extrapolation
Figure A.44 Wavenumber spectrum of a circular Fourier reconstructed spatial kz convolution operator (5x5).
the data, a new term is added to the series expansion. The scheme given in figure
A.43 is more sensitive to numerical errors than the Chebychev recursion scheme but
if the number of terms remains small this causes no stability problems. It is interesting to note that an exact analytical expression for H1 (x, y) (H1 ⇐⇒H̃1 = kz ) can
be derived (see Berkhout (1982), appendix E). However, a weighted least-squares
version yields a shorter operator.
The factors âm of the series expansion in figure A.43 are obtained by a least-squares
Figure A.45 Wavenumber spectrum of the approximated phase shift operator with 9 terms in the
series expansion, 128 × 128 kx , ky points with c = 1000 m/s, f = 25 Hz, ∆x =
∆y = ∆z = 10 m and a maximum angle of interest of 65◦ .
Appendix A: Operator optimization
-500
-250
x-position [m]
0
250
-500
500
-250
247
0
250
500
0
-500
250
y-position [m]
-250
0
500
0
250
250
500
z = 220
500
x=0
x=y
10x(5x5) kz series operator
Figure A.46 Pulse responses obtained with the series expansion in kz with a small basis operator
(5x5) with 10 terms in the series expansion. Note the artifacts at the higher angles.
optimization method with respect to the wavenumber spectrum of the optimized H1
operator. The circular Fourier reconstruction of the kz operator results in a circular
short spatial convolution operator, the spectrum of an operator is shown in figure
A.44 for a 5 × 5 operator. Using this 2-D convolution operator the computational
effort can be reduced in comparison with the direct method if the number of terms
in the series expansion remains small. Starting with 4 terms in the expansion it
was observed that by adding more terms in the series expansion both the amplitude
and phase error decrease. However, beyond 9 terms the error remains the same and
increasing the number of terms does not improve the result. In that case a better
approximation is not possible with the optimized basis function of H1 . For a better
approximation result a better approximation to kz must be chosen.
In figure A.45 the wavenumber spectrum of the approximated phase shift operator
is shown for a nine order series expansion with optimized coefficients for the approximated kz operator with a spatial size of 5 × 5 points. Outside the band of
interest (αmax = 65◦ ) the exact kz operator (which is used as object function in the
optimization) is tapered to zero. Note that with only nine terms there is already
a good match with the analytical spectrum. In figure A.46 the pulse response is
shown for the single series expansion method. It is observed that convolution with
the short basis function of 5×5 points gives an image with a non-circular depth slice.
Increasing the number of terms in the series expansion from 10 to 15 will reduce the
artifacts around the event, but the non-circular behavior remains the same. A longer
basis function with 7 × 7 points in the spatial domain gives a better circular slice,
but requires much more computation time.
248
A.2 2-Dimensional operators for 3-dimensional extrapolation
Chebychev expansion in kz with L∞ -norm
Using an optimization method which makes use of the L∞ norm and the Chebychev
recursion structure for the implementation of the convolution may improve the result
of the kz expansion. Optimization for the series expansion terms with the L∞ norm
can be done by reducing the polynomial synthesis to symmetrical spectral synthesis
with the aid of a simple transformation (Rabiner and Gold, 1975, page 151). In
this optimization only the extreme values of the wavenumber spectrum of the H1
operator are used and not, as with the L2 optimization, the whole spectrum of the
H1 operator. The transformation from polynomials to spectral synthesis reduces the
2-D optimization problem to a 1-D optimization problem which can be solved with
the Remez exchange algorithm in a fast way.
In figure A.47 the impulse response is shown for a 5 point basis convolution operator
with 10 terms in the expansion. As was observed before the basis operator is not
accurate enough to define the circular shape properly. A 7 point operator with the
same number of terms will gives a better result. The difference with L2 optimization
is that due to the equiripple character of the L∞ solution the error is smeared over
the whole wavenumber domain, while in the L2 optimization the biggest error occurs
at the edges of the domain of interest (defined by k = ωc and α) which gives the
artifacts as shown in figure A.46. Note that more terms in the series expansion with
the same basis operator will reduces the artifacts but it will not improve the circular
shape.
-500
-250
x-position [m]
0
250
-500
500
-250
0
250
500
0
-500
250
y-position [m]
-250
0
500
0
250
250
500
z = 220
500
x=0
x=y
10x(5x5) kz in Chebychev recursion scheme
Figure A.47 Depth image of an impulse response obtained with the Chebychev recursion scheme
and a L∞ optimization for the coefficients in the expansion in kz , a short basis operator
(5x5) with 10 terms is used.
Appendix A: Operator optimization
249
Error analysis
In table A.10 the errors are given for the expansion in kz with the direct implementation and the Chebychev recursion scheme. From the results in the table the
following remarks can be made;
(1) The expansion in kz gives large ε2 errors for most frequencies. So the operator
is not very accurate which is observed in the artifacts in the impulse responses.
(2) Increasing the number of terms with the same size of the basis operator does not
improve the result. Increasing the size of the basis operator with the same number
of terms gives an improvement. This means that the approximation to the basis
operator is the most important factor in the performance of the operator.
(3) The εp error indicates that the used basis operator of size 5 × 5 in the single
series expansion does not have a good circular shape. The larger operator of size
7 × 7 gives an improvement but is still inaccurate.
(4) Optimization of the series coefficients with the L∞ norm and the use of the
Chebychev recursion scheme gives a better result than L2 optimization and a direct
recursion scheme.
The shortest accurate operator as function of the maximum angle of interest is not
worked out for the expansion in kz because this method cannot be designed accurate
enough within a reasonable computation effort.
size
Series
ε2
5 Hz
εa∞
εp
ε2
20 Hz
εa∞
εp
ε2
40 Hz
εa∞
εp
10x(5x5)
15x(5x5)
10x(7x7)
15x(7x7)
1.8e-3
7.4e-4
5.8e-4
4.7e-4
1.8e-3
6.3e-4
1.2e-3
1.2e-3
4.3e-4
1.4e-4
1.1e-4
6.5e-5
1.5e-2
1.5e-2
7.6e-3
7.6e-3
3.6e-3
3.9e-3
1.8e-3
3.8e-3
7.5e-3
7.6e-3
3.5e-3
3.5e-3
2.0e-2
2.0e-2
1.2e-2
1.2e-2
5.6e-3
5.2e-3
1.9e-3
4.0e-3
7.5e-3
7.1e-3
8.9e-3
8.6e-3
Chebychev
10x(5x5)
15x(5x5)
10x(7x7)
15x(7x7)
1.9e-2
2.0e-2
9.2e-2
9.2e-3
5 Hz
2.1e-3
2.2e-3
2.5e-3
2.1e-3
2.8e-3
2.7e-3
2.6e-3
2.7e-3
2.1e-2
2.1e-2
3.6e-2
3.6e-2
20 Hz
2.8e-3
4.3e-3
1.1e-2
1.5e-2
9.1e-3
8.9e-3
1.4e-2
1.4e-2
3.5e-2
3.5e-2
1.4e-2
1.3e-2
40 Hz
2.5e-3
2.5e-3
2.7e-3
2.4e-3
5.2e-2
5.3e-2
2.1e-2
2.1e-2
Table A.10 Errors in the extrapolation operators for the direct series expansion in kz and a
Chebychev recursion scheme in kz . The maximum angle of interest is 60◦ .
250
A.2.4
A.2 2-Dimensional operators for 3-dimensional extrapolation
Expansion in kx2 + ky2
By using an additional series expansion, i.e. kz is expanded in terms of kx2 +ky2 (= kr2 )
(see equation (A.68)) ,
kz = k
p
k 2 (k 2 )2 (k 2 )3 5 (kr2 )4 7 (kr2 )5
6
−
+ O((kr2 ) ) (A.68)
1 − kr2 ≈ k (1 − r − r − r −
2
8
16
128
256
there is an extra advantage (chapter 10 Berkhout, 1982, see). The basic spatial
convolution operators are reduced to the simple 1-D convolution operators: d2 (x)
and d2 (y). The double series expansion in kx2 + ky2 is given in equation (A.69)
exp (jkz ∆z) ≈ 1−
j∆z
2
2k (kx
j∆z
8k3 (1
j∆z
16k5 (1
+ ky2 ) −
− jk∆z)(kx2 + ky2 )2 −
(jk∆z)2
)(kx2
3
O((kx2 + ky2 )4 )
− jk∆z +
+ ky2 )3 +
(A.69)
where the terms of the series expansion are derived from the Taylor series, but this
is not an optimum choice (Hoff, 1995). This same expansion can also be regarded as
an approximation to the cosine terms in equation (A.46). So there are two different
ways to look at the double series expansion
Ỹ0 (kx , ky ) = exp (−jkz ∆z)
≈
≈
≈
M
X
(A.70)
am Tm (cos (kr ))
(A.71)
bm Tm (kr )
(A.72)
cm krm
(A.73)
m=0
M
X
m=0
M
X
m=0
where Tm is a Chebychev polynomial of the m’th order. First the series of equation
(A.73) is discussed and next the Chebychev recursion of equation (A.72) is discussed.
Series expansion in kx2 + ky2 with L2 -norm
The same techniques as discussed in the section with the series expansion in kz can be
used again, the spatial versions of kx2 and ky2 , i.e. d2 (x) and d2 (y), are determined by
a weighted least-squares process. In figure A.48 two wavenumber spectra are shown
for short spatial convolution operators, with operator lengths of 5 and 7 points, which
represent the second order differentiation. Note that for these short operator the
approximation to the exact function within the band of interest is within a reasonable
Appendix A: Operator optimization
0.10
0.10
7 point d2 (x) operator
0.08
amplitude
251
7 point d2 (x) operator
0.08
0.06
0.06
0.04
0.04
0.02
0.02
0
0
0
0.1
0.2
0.3
0.4
0.5
0
0.1
cycles
0.2
0.3
0.4
0.5
cycles
Figure A.48 Spectrum of two differentiation operator for 5 and 7 points. The maximum wavenumber
= 0.25.
of interest is given by kxmax = 2π25
1000
error. The convolution scheme is given in figure A.49 where d2 (x, y) stands for the
spatial Laplacian operator d2 (x) + d2 (y). The factors b̂m of the series expansion
in figure A.49 are obtained by a weighted least-squares optimization method with
respect to the wavenumber spectrum of the optimized kx2 + ky2 operator. The length
of the 1-D convolution operators depends on the maximum wavenumber of interest.
Hoff (1995) showed that the value of the coefficients in the series expansion grows
rapidly with increasing order. For the higher order terms values in the order of 1e19
are normal. This means that after a certain number of terms the accuracy cannot
be improved any further due to the limit of the floating point representation in the
computer. This indicates also that adding higher order terms will not improve the
result significant.
The WLSQ optimization scheme proposed in this section uses the whole wavenumber
spectrum to obtain the series coefficients. In figure A.50 the impulse response is
shown where 15 coefficients are used in the series expansion and different operator
lengths are used for the short convolution operators. For the small wavenumbers an
operator length of 3 points is used for the representation of the differential operator,
for larger wavenumber values the length increases with 2 points up to 7 point for
the largest wavenumbers. The impulse response in figure A.50 shows artifacts at the
Pi
d2(x,y)
b̂0
d2(x,y)
b̂1
+
d2(x,y)
b̂2
b̂
+
M
1
+
b̂
M
+
Pi+1
Figure A.49 Series expansion in terms of kx2 + ky2 = L̃. The d2 (x, y) boxes represent the two 1-D
convolutions with the optimized Laplacian operator, the b̂m represent the optimized
coefficients in the series expansion.
252
A.2 2-Dimensional operators for 3-dimensional extrapolation
-500
-250
x-position [m]
0
250
-500
500
-250
0
250
500
0
-500
250
y-position [m]
-250
0
500
0
250
250
500
z = 220
500
x=0
x=y
15x(3,5,7) with WLSQ coefficients
Figure A.50 Depth images of pulse responses obtained with the series expansion in kx2 + ky2 .with a
basis operator with optimized operator lengths with 15 terms in the series expansion
(example made by Jochum Hoff).
higher angles which are due to edge effects of the used WLSQ method. In WLSQ
design the edges of the domain of interest contain relative large error peaks. It may
therefore be better to use an L∞ norm in the design of the series coefficients.
Chebychev expansion in kx2 + ky2 with L∞ -norm
Soubaras (1996) used the same type of expansion in kx2 + ky2 , but in his method the
optimization technique for both the terms in the series expansion and the convolution
operators is the Remez exchange algorithm with the L∞ norm. The advantage of the
Chebychev recursion scheme, given in figure A.51 over the series expansion, which
was discussed in the previous subsection, is that the coefficients in the Chebychev
expansion are less sensitive to numerical errors. Sollid and Arntsen (1994) used also
−
Pi
L(x,y)
2L(x,y)
−
2L(x,y)
2L(x,y)
−
B0
2B1
+
2B2
+
2BM
+
Pi+1
Figure A.51 Expansion in terms of kx2 + ky2 = L̃. The d2 (x, y) boxes represent the two 1-D convolutions with the optimized L operator, the Bm represent the optimized coefficients in
the series expansion.
Appendix A: Operator optimization
-500
-250
x-position [m]
0
250
-500
500
253
-250
0
250
500
-250
0
250
500
0
-500
250
y-position [m]
-250
0
500
0
250
250
500
z = 220
500
x=0
x=y
a. 12x(7+7) expansion in kr
-500
-250
x-position [m]
0
250
-500
500
0
-500
250
y-position [m]
-250
0
500
0
250
250
500
z = 220
500
x=0
x=y
b. 15x(7+7) expansion in kr
Figure A.52 Depth images of pulse responses obtained with the series expansion in kx2 +ky2 with L∞
optimization; a) shows a 7 point 1D convolution operator with 12 terms in the series
expansion and b) with the same basis operator and 15 terms in the series expansion.
Note that more terms in the series expansion gives a better result.
the Chebychev recursion scheme but the optimization is done with respect to the L2
norm and non-linear methods were used to obtain the coefficients of the expansion.
Optimization of the differentiation operators with the L∞ norm gives equiripple operators. Optimization for the series expansion terms with the L∞ norm can be done
by reducing the polynomial synthesis to symmetrical spectral synthesis with the aid
of a simple transformation. In this transformation only the extreme values of the
254
A.2 2-Dimensional operators for 3-dimensional extrapolation
wavenumber spectrum of the differentiation operator are used in the optimization
and not, as with the L2 optimization the whole spectrum of the differentiation operator. The transformation from polynomials to spectral synthesis reduces the 2-D
optimization problem to a 1-D optimization problem which can be solved with the
Remez exchange algorithm in a fast way. In the extrapolation scheme the differentiation operator remains the same for all frequencies. However, the terms in the series
expansion are calculated for every frequency.
In figure A.52 two pulse responses are shown for a 7+7 point operator with 12 and 15
terms in the expansion. The pulse response obtained with 12 terms in the recursion
scheme is shown in figure A.52a and has a circular response with only small artifacts.
Using 15 terms in the expansion gives an even better result with less artifacts.
Error analysis
In table A.11 gives the errors for the series expansion in kx2 + ky2 implemented in a
direct recursion scheme with L2 optimization and the Chebychev recursion scheme
with L∞ optimization. From the results in the table the following remarks are made;
(1) In both optimization methods the number of terms is less essential to the accuracy
than the length of the d2 operator. Increasing the number of terms improves the
result only a little, while increasing the length of the d2 operator gives a significant
improvement on the result.
(2) The artifacts present in the impulse response with the series expansion scheme
are due to instabilities at the higher angles. This effect is possibly due to the use of
the L2 norm optimization. In the L∞ norm optimization unstable error peaks are
not likely to occur. The artifacts can be removed with an additional optimization
step in the L2 optimization (Hoff, 1995).
(3) Although the ε2 errors for the lower frequencies in the L∞ norm optimization
are larger than the errors in the L2 norm optimization the impulse response contains
less artifacts at the higher angles.
size
Direct
ε2
5 Hz
εa∞
ε2
20 Hz
εa∞
ε
ε
10x(3,5,7)
12x(3,5,7)
15x(3,5,7)
5.8e-3
6.1e-3
6.1e-3
1.9e-3
2.8e-3
3.5e-3
1.8e-3
1.9e-3
1.9e-3
3.2e-3
3.0e-3
2.4e-3
3.7e-3
3.2e-3
3.2e-3
Chebychev
10x(7+7)
12x(7+7)
15x(7+7)
4.0e-2
4.0e-2
4.0e-2
5 Hz
3.7e-3
7.6e-3
1.2e-2
2.3e-3
2.1e-3
2.4e-3
7.1e-3
7.1e-3
6.8e-3
20 Hz
3.0e-3
5.0e-3
5.8e-3
p
ε2
40 Hz
εa∞
εp
4.6e-3
4.1e-3
3.3e-3
3.4e-3
3.1e-3
3.0e-3
3.4e-3
3.4e-3
3.9e-3
6.4e-3
5.1e-3
4.8e-3
4.5e-3
4.4e-3
3.4e-3
4.3e-3
4.1e-3
4.1e-3
40 Hz
1.9e-3
4.9e-3
6.5e-3
6.5e-3
5.8e-3
5.0e-3
p
Table A.11 Errors in the extrapolation operators for the direct series expansion in kx2 + ky2 and a
Chebychev recursion scheme in kx2 + ky2 . The maximum angle of interest is 60◦ .
Appendix A: Operator optimization
255
0.8
-500
-250
x-position [m]
0
250
500
0.6
error
-500
0.2
y-position [m]
-250
0
0◦ 15◦
0
0
250
250
500
z = 220
500
a. Series expansion with L2
-500
-250
x-position [m]
0
250
30◦
45◦
60◦
65◦
30◦
45◦
60◦
65◦
x=y
0.8
500
0.6
error
-500
0.4
0.2
-250
y-position [m]
0.4
0
0◦ 15◦
0
0
250
250
500
z = 220
500
x=y
b. Chebychev expansion with L∞
Figure A.53 Error in the impulse response of the Laplacian expansion in kx2 + ky2 for a series
expansion and a Chebychev recursion scheme.
The difference with the reference error in the spatial domain is given in figure A.53.
In this figure the series expansion method with 15x(3,5,7) terms and L2 optimization
can be compared with the Chebychev recursion scheme of 15x(7+7) terms and L∞
optimization. From this comparison we see that the errors in the L∞ optimization
are smeared out over the whole wavenumber range while the L2 optimization has
an error peak. The amplitude and phase accuracy is better for the L2 optimization,
but it suffers from artifacts at the higher angles. Choosing larger basis convolution
operators to represent the differentiation will solve this problem.
256
A.2 2-Dimensional operators for 3-dimensional extrapolation
Operator
angle
size
15
30
45
60
5x(7+7)
5x(7+7)
7x(7+7)
9x(7+7)
ε2
20 Hz
εa∞
εp
ε̄2
Average
ε̄a∞
ε̄p
8.6e-3
3.0e-3
5.0e-3
8.3e-3
5.1e-3
1.9e-3
4.5e-3
1.8e-3
1.7e-4
3.9e-3
2.0e-3
6.3e-2
9.6e-3
1.2e-2
1.5e-2
1.7e-2
9.7e-3
6.1e-3
5.8e-3
2.3e-3
5.0e-4
3.7e-3
1.0e-2
4.0e-2
Table A.12 Optimum operators which are accurate up to a maximum angle of interest.
In table A.12 the shortest accurate operator in L∞ optimization is given as function
of the maximum angle of interest. The ε2 error is the most sensitive error in the
L∞ optimization. For 15◦ and 30◦ angle the same number of terms must be used,
a lower number of terms leads to unacceptable ε2 errors for the low and middle
frequencies. The optimum operator size can be found by trying to make the ε2
as small as possible by choosing the number of terms high. The optimum number
of terms is then found by lowering the number of terms until the ε2 is changing
significant. The high average errors are due to the high errors at low frequencies.
For a maximum angle of 75◦ it was not possible to obtain stable operators within
the current implementation.
A.2.5
Computation times
The computation times of the different 3-D extrapolation methods in the spacefrequency domain is given in table A.13. The given time represent 55 recursive
depth steps for one frequency (20 Hz.) with c = 1000 m/s, ∆x = ∆y = ∆z = 10 m
on a x,y grid of 111 × 111 samples wide. All routines which are used are written in C
and Fortran and are translated with the same type of compiler options without using
options for parallel computation (see Table A.14). However, parallel processing is
easily implemented on the main frequency loop in the extrapolation algorithm. On
the Convex (C-220) the -O2 option is used for vectorization of the loops. It was
not possible to vectorize the C-code with specific compiler directives, therefore the
convolution schemes were written in Fortran code which vectorize well. Note that
in general the Fortran compilers are better in optimization than the C compilers.
In the direct implementation of the 2-D convolution the even symmetry in the convolution operator is used. This implementation is designed to work fast on a Vector
computer. In the implementation of the McClellan transformation and the kz expansion the circular symmetry in the basis operators is used by first adding the
common terms to each other and then multiplication with the appropriate operator
point. This reduces the number of multiplications with a factor 8 in comparison
with a full convolution. The computation times given in table A.13 are real-time
computation times measured during the calculation. The time needed to calculate
the operators is not included in this time.
Appendix A: Operator optimization
257
Due to the use of the even symmetry in the operator the computation time for the
direct convolution is a real challenge for the other methods (Note that the direct
scheme can be made even faster when ∆x = ∆y and the circular symmetry is used).
The first and second order McClellan implementations and the series expansion in
cos (kr ∆x) are the fastest algorithms on all machines. The series expansion in kz
and in kx2 + ky2 are comparable with the McClellan transformation. The difference
between the kx2 +ky2 expansion in L2 and L∞ is that in the L2 scheme the direct series
expansion is used and for the L∞ the Chebychev recursion scheme is used. From the
Machine
19x19
Direct
25x25
31x31
McClellan 1
10x9 13x9 16x9
SUN C
SUN F
Convex F
DEC C
DEC F
HP C
HP F
63.5
52.6
34.7
28.3
11.5
18.6
6.8
104.6
88.0
61.9
46.9
19.9
30.0
10.6
157.3
144.8
85.9
71.6
28.0
44.7
15.1
24.5
20.3
12.4
9.4
5.6
10.0
6.6
Machine
SUN C
SUN F
Convex F
DEC C
DEC F
HP C
HP F
McClellan 2+
10x25
13x25
16x25
41.9
36.8
33.4
14.6
7.5
23.3
15.1
55.0
48.0
45.5
18.5
9.6
26.4
18.2
67.2
58.3
54.1
24.3
11.8
30.4
22.0
Machine
10x(5x5)
kz
15x(5x5)
10x(7x7)
SUN C
SUN F
Convex F
DEC C
DEC F
HP C
HP F
36.3
33.2
31.5
11.8
6.2
12.8
9.5
54.3
49.0
49.6
17.6
9.1
18.8
14.7
61.2
57.2
71.1
26.3
8.5
18.5
20.1
31.1
27.0
17.4
13.7
7.2
13.2
8.6
38.7
31.7
21.6
15.7
8.8
16.5
10.6
cos (kr ∆x) ×9
10x9 13x9 16x9
18.0
16.8
11.6
7.0
4.4
7.9
4.0
23.5
21.1
16.2
9.3
5.6
10.3
5.1
29.2
25.7
18.5
10.8
6.7
12.3
6.2
McClellan 2
10x17 13x17 16x17
33.4
28.8
21.1
13.4
6.7
20.5
19.2
44.8
39.0
27.6
17.8
8.6
23.5
20.8
55.3
47.8
33.5
21.5
10.5
27.7
24.5
cos (kr ∆x) ×25
10x25 13x25 16x25
35.0
32.6
34.0
12.1
6.2
13.4
12.2
45.9
42.9
42.3
15.6
7.9
17.0
14.6
59.9
54.3
53.9
19.2
9.6
20.8
17.4
kx2 + ky2 L2
12x5 12x7 15x7
kx2 + ky2 L∞
12x5
12x7
15x7
25.7
21.2
15.1
8.6
5.6
15.9
13.5
33.4
27.9
19.0
13.6
7.0
14.6
11.8
32.8
26.2
19.7
9.6
5.6
21.4
20.0
41.2
35.5
24.3
11.9
6.8
27.4
25.3
41.9
32.0
24.2
15.2
7.7
18.1
15.8
51.7
39.9
30.0
19.2
9.8
22.3
18.3
Table A.13 2-D convolution computation time (in seconds) for one frequency on different machines
for different operator sizes and extrapolation methods.
258
A.2 2-Dimensional operators for 3-dimensional extrapolation
table it is clear that the hardware design of the computer system can be optimum for
some specific implementation. For example the HP has a good performance on the
direct scheme and less on the McClellan schemes. The DEC has a good performance
on all expansion schemes.
Which scheme is preferred depends also on the desired accuracy of the result. If one
uses the extrapolation only to get a first idea of the subsurface, or to estimate the
macro model, a first order McClellan scheme can be used. For a higher accuracy the
direct convolution or a series expansion method with a high number of terms can be
used.
In table A.14 a detailed specification for the different machines is given. On the SUN
the gcc compiler is used because it produces faster code than the standard cc compiler
delivered by SUN. Note that for the optimization options only the most common
used options are chosen, it may therefore be possible that by choosing another option
the scheme will perform better as described in this Appendix (suggestions for better
options are welcome).
Machine
type
RAM (Mb)
bits
C
F
C-opt
F-opt
SUN
Convex
DEC
HP
10/514
C-220
3000-500
9000-735
256
256
96
144
32
32
64
32
gcc
cc
cc
cc
f77
fc
f77
fort77
-O2
-O2
-O2
+O4
-O2
-O2
-O3
+O4
Table A.14 Specification for the used machines. Note that the optimization is done for one CPU and
parallel processing is not used.
A.2.6
Concluding remarks
The 3-D extrapolation algorithm that is used in recursive depth migration can be
implemented in several ways. In this section the direct method, the McClellan transformation and three series expansion methods have been discussed. For the direct
method a 2-dimensional convolution operator is needed. The proposed weighted
least-squares optimization method is an efficient procedure which gives stable and
accurate convolution operators (Thorbecke and Rietveld, 1994). This method can be
further improved by a second optimization step; for example, the Lawson algorithm
(Rice and Usow, 1968), which will adjust the weight function in such a way that
after several steps it will converge to a Chebychev-norm solution see for example Algazi et al. (1986). The Fourier reconstruction method is a fast and simple method
to obtain 2-D circular convolution operators but must be improved further to give
accurate results.
The McClellan scheme which makes use of the 1-D optimized operator coefficients
Appendix A: Operator optimization
259
is attractive with respect to the computation effort and by using optimized McClellan factors the accuracy for the higher angles can be improved significantly without
much effort. Using a series expansion of the phase shift operator also reduces the
computation time in comparison with a direct 2-dimensional convolution. Two expansions were discussed in this section one in kz and one in kx2 +ky2 . The kz expansion
is not recommended because it is difficult to optimize the parameters used in the
optimization.
All these different approaches to the phase shift operator can be summarized in the
following equations
Ỹ0 (kx , ky ) = exp (−jkz ∆z)
≈
Ỹ0 (kx , ky ) ≈
≈
Ỹ0 (kx , ky ) ≈
≈
Ỹ0 (kx , ky ) ≈
≈
M X
N
X
Ymn cos (kx m∆x) cos (ky n∆y)
(A.74)
(A.75)
m=0 n=0
M
X
m=0
M
X
m=0
M
X
q
Ym Tm (cos ( kx2 + ky2 ∆x))
(A.76)
q
am cosm ( kx2 + ky2 ∆x)
(A.77)
Bm Tm (kx2 + ky2 )
(A.78)
bm (kx2 + ky2 )m
(A.79)
Cm Tm (kz )
(A.80)
cm kzm
(A.81)
m=0
M
X
m=0
M
X
m=0
M
X
m=0
with
Q
P X
q
X
cpq cos (pkx ) cos (qky )
cos ( kx2 + ky2 ) ≈
(A.82)
p=0 q=0
q
q
cos ( kx2 + ky2 ) ≈ a0 + b1 kx2 + ky2
(A.83)
Equation (A.75) represents the direct method, equation (A.76) combined with equation (A.82) is the McClellan approach with the Chebychev recursion scheme. Equation (A.77) represents the series expansion in cos (kr ∆x) with a dependent optimization between the series coefficients and the approximation to cos (kr ∆x). Equation
(A.78) combined with equation (A.83) represents the expansion in kx2 + ky2 with the
260
A.2 2-Dimensional operators for 3-dimensional extrapolation
Chebychev recursion scheme. Equation (A.79) is the series expansion in kx2 + ky2 , the
use of this series expansion in recursive migration was already proposed by Berkhout
(1982). Equation (A.80) represents the expansion in kz in a Chebychev recursion
scheme and Equation (A.80) represents the series expansion in kz .
From an efficiency point of view the expansion in kx2 +ky2 and the use of the McClellan
transformation are the most interesting schemes. The approximation to cos (kr ∆x)
used in the McClellan transformation can be done with many different methods.
Crucial in the performance of the extrapolation operator is that the coefficients in
the expansion (Chebychev or series) are optimized by using this approximation to
cos (kr ∆x). The big advantage of the expansion in kx2 +ky2 is that short 1-dimensional
convolution operators can be used instead of the small 2-dimensional convolution
operators used in the McClellan schemes. A disadvantage of these schemes is that it
is not possible to write the algorithms in computer ’friendly’ way due to the recursive
structure in the scheme. This fact is displayed in the computation times given
in table A.13. The most accurate extrapolation is the direct convolution scheme.
Another advantage of the direct scheme is that the algorithm can be designed in
an efficient way. A disadvantage of the direct scheme is the intensive computation
needed to compute the 2-dimensional convolution operators.
Method
accuracy
stable
circ
operator
simple
vector
scalar
Direct
McC1
McC2
McC2+
cos (kr ∆x)
kz
kx2 + ky2 L2
kx2 + ky2 L∞
++
+
+/++
+
++
++
+
+
+
++
+
++
+
+
+
+
+
+
++
++
+
+
-
+
+
+
+
+
++
++
+
++
+
+
Table A.15 Comparison of the different extrapolation methods with respect to computation effort
and stability. Note that in the McClellan schemes the optimized McClellan factors are
used.
In table A.15 a simplified summary is given for the different extrapolation schemes
used in the section. In table A.15 the different columns have the following meaning:
• accuracy: the average ε2 error over the whole frequency range.
• stable: the stability of the method over the whole frequency range for all wavenumbers (εa∞ error). A o means that some wavenumber components can become unstable.
• circ: the circularity of the impulse response (εp error). The McClellan in
Chebychev expansion and the kz scheme have problems with the circularity.
• operator: the effort to compute all the coefficients which are needed in the convo-
Appendix A: Operator optimization
261
lution scheme. For example in the direct method a 2-D convolution operator must
computed, in the McClellan scheme a 1-D convolution operator and the (optimized)
McClellan factors are needed. In the table ++ means a minimum computation effort
to compute the coefficients. Note that the operator coefficients can be calculated in
advance and stored in an operator table.
• simple: the simplicity of the implementation of the convolution. The recursive
schemes require more complex algorithms, so the compilers have to be good in optimization to make these schemes fast. In the recursive schemes it is difficult to make
the program faster by changing the algorithm in a more computer ’friendly’ way. A
direct convolution requires more multiplications and additions but the algorithm can
be made very efficient. This fact explains the fast computation time of the direct
scheme in comparison with the other schemes (see Appendix A).
• vector: the performance of the scheme on a vector computer. The direct scheme
is the only scheme which can be implemented in a vector efficient way.
• scalar: the performance of the scheme on a modern scalar computer. Note that
some scalar computers may have an architecture which can be more advantageous
for some implementations.
In conclusion; taking into account the computation time of the different methods,
the simplicity of the algorithms and most important the accuracy of the result then
the direct method (A.75) is the best method for 3-D extrapolation. The 2-D
convolution operators should be stored in an efficient way, by using the symmetry
of the operator (one octant need be stored only), in an operator table that can
be calculated in advance. If a series expansion version is used the expansion in
cos (kr ∆x) is preferred, Chebychev recursion is not an advantage.
262
A.2 2-Dimensional operators for 3-dimensional extrapolation
Appendix B
Matrix notation
In this appendix the matrix notation of seismic data, first introduced by Berkhout
(1982) and used in this thesis, will be explained in more detail. The matrix notation
takes the discrete sampling, in both space and time, of seismic data into account.
A consequence of this discrete representation of seismic data is that all wave theory
based operations are carried out as discrete summations and multiplications in the
computer. Nice features of the matrix notation are its simplicity for the description
of (one-way) wavefield operators, for example propagation and reflection, and the
close relation with the actual implementation in a computer algorithm. The matrix
notation will be introduced for 2-dimensional and 3-dimensional seismic data.
B.1
2-Dimensional wavefields
As the earth is considered as a time-invariant medium the seismic problem can be
described in frequency components which are independent from each other. Consider
a 2-dimensional wavefield, measured for only one component (e.g. pressure) at the
surface at a constant depth level zr and originating from a source at xs , described
by
p(xr , xs , zr , t),
(B.1)
Then the temporal Fourier transform of this wavefield is defined as
p(xr , xs , zr , t)@ >> Ft > P (xr , xs , zr , ω),
(B.2)
where Ft stands for the Fourier transformation from time to frequency. From Fourier
theory it is known that the negative frequency components of the Fourier transformation of a real signal, are by definition equal to the complex conjugate of the positive
frequency components. As the measured wavefield is a causal and real signal, the
positive frequency components describe the wavefield completely. Therefore in the
discretized version of equation (B.2) only the positive frequency components have
264
B.1 2-Dimensional wavefields
to be taken into account. The discretized version for one (positive) frequency component reads the so called data vector
T
P(zr ) = [P (∆xr , zr , xs ), . . . P (m∆xr , zr , xs ) . . . P (M ∆xr , zr , xs )] ,
where the dependency on frequency is omitted for notational convenience. Combining several monochromatic data vectors for different seismic experiments into one
matrix gives (one seismic experiment: fixed xs )
P (∆xr , zr , ∆xs )
6 P (2∆xr , zr , ∆xs )
6
6
..
6
.
6
P(zr ) = 6
6 P (m∆xr , zr , ∆xs )
6
6
..
4
.
P (M ∆xr , zr , ∆xs )
2
...
...
..
.
...
...
...
. . . P (∆xr , zr , n∆xs )
. . . P (2∆xr , zr , n∆xs )
...
...
...
. . . P (m∆xr , zr , n∆xs )
..
.
. . . P (M ∆xr , zr , n∆xs )
...
...
...
...
3
P (∆xr , zr , N ∆xs )
P (2∆xr , zr , N ∆xs ) 7
7
7
..
7
.
7
7,
P (m∆xr , zr , N ∆xs ) 7
7
7
..
5
.
P (M ∆xr , zr , N ∆xs )
which is known as the data matrix and where the variable zr indicates the depth level
z = zr to which the matrix is related to. Each element Pm,n = P (m∆xr , zr , n∆xs ) in
the data matrix corresponds to a fixed receiver coordinate and a fixed lateral source
coordinate. One column of the data matrix represents one monochromatic seismic
experiment (common source gather) and one row represents one common receiver
gather. The diagonal elements of the data matrix contain common offset data with
the zero offset data on the main diagonal, the anti-diagonal elements contain common
midpoint data. In figure B.1 a pictorial representation of the construction of the
data matrix is shown. A shot record measured at the surface is transformed to
the frequency-domain. One frequency component of this transformed shot record is
placed in the data matrix as a column vector at its source position.
For a fixed spread configuration the data matrix is completely filled. However, for
true seismic acquisition patterns the data matrix is generally differently filled due
to the used acquisition pattern in the field. For example, missing near offsets can
be recognized as a band matrix with the band around the diagonal filled with zeros,
on the other hand one sided marine data has an upper or lower matrix structure.
With the introduced matrix notation a lateral convolution operator (e.g. wavefield
propagation) on a single frequency component of the wavefield is equivalent to a
matrix-vector multiplication. As an example of a matrix operator working on the
data matrix the one-way propagation matrix is discussed in more detail in the next
subsection.
Propagation operators
The W matrices, defined in the forward model discussed in chapter 3, represent the
operation of propagation of a wavefield from one depth level to another depth level
Matrix notation
265
xr
xs
t
x x x x
xr
xs
x x x x
ωi
FFT
ω
x
x
x
ωi
x
x
x
x
x
P(zr )
common
receiver
gather
common
shot gather
common
offset gather
x
x
x
zero offset
gather
Figure B.1 The definition of the data matrix P(zr ). The matrix P(zr ) consists of one monochromatic seismic experiment combined for different shot positions into one matrix. Within
the data matrix several well known seismic data gathers can be recognized.
as shown in figure B.2 and which is mathematically expressed by
P + (zm ) = W+ (zm , z0 )P + (z0 ).
(B.3)
One column vector of the data matrix (which represent the receivers from one shot)
is extrapolated from depth level z0 to depth level zm by a spatial convolution process. The convolution operators, obtained from one of the procedures described in
appendix A, are placed as row vectors in the propagation matrix W+ . Every single
operator describes the propagation from one depth level to another depth level for
the same lateral position. In figure B.2 it can be seen how one operator is stored
as a row vector in the propagation matrix W+ . For a 1-dimensional medium all the
operators are the same for all lateral positions, meaning that the propagation matrix
W+ has a Toeplitz structure.
Note the similarity of the propagation matrix W+ with the data matrix P+ (z0 ).
Note also an important difference. Opposed to the data matrix, which contains
physical wavefields obtained from acoustic experiments, the propagation matrix con-
266
B.2 3-Dimensional wavefields
P + (z0 )
x x x x x x x x x x x x x
z0
W+ (zm , z0 )
P + (zm )
zm
=
Figure B.2 Pictorial presentation of the propagation matrix W+ ; one extrapolation (convolution)
operator for one lateral position is stored as a row vector in the propagation matrix. For
a 1-dimensional medium W+ is a Toeplitz matrix.
tains mathematical wavefields obtained numerically from the macro model and one
of the optimization methods discussed in appendix A.
B.2
3-Dimensional wavefields
The 3-dimensional measurements are combined in a 2-dimensional matrix, similar
to the 2-dimensional case. Again all traces are transformed to the frequency domain
and data matrices per Fourier component are constructed. The data, for all receivers
in the x − y plane, of one shot record position are stored into one column of the
data matrix as shown in figure B.3. Within this matrix notation one column still
corresponds with a common source gather and one row with a common receiver
gather. The matrix multiplications describe now 2-dimensional spatial convolution
in the x−y direction. Note that diagonals do not contain common offset information
anymore, except for the main diagonal (Verschuur, 1991).
Matrix notation
y
z = z0
x
267
seismic
shot record
xs
xr
seismic
line
ysource
yreceiver
Figure B.3 A data matrix for a 3D acquisition on a grid in the x − y plane consists of a 2D sub
matrix for each y−source coordinate and for each receiver y−coordinate. One column
still describes a multi streamer shot record (in this example 5 streamers are shown)
and one row describes a 3D common receiver gather (figure used with courtesy of Eric
Verschuur).
268
B.2 3-Dimensional wavefields
Appendix C
Algorithms
CFP technology involves two focusing steps. The first focusing step transforms shot
records to CFP gathers and the second focusing step transforms CFP gathers to
image traces. In CFP technology preprocessing and velocity analysis occurs between
the two focusing steps. In this appendix the calculation algorithms of the two
focusing steps and the analysis modules are discussed. The difference between time
and frequency processing is also explained and the numerical implementation of the
time-domain scheme is discussed in more detail.
C.1
Processing flow
The general block diagram for CFP processing is shown in figure C.1, where the
emphasis is put on the data-control during the processing. To construct the CFPgather from the shot records an initial focusing operator is needed. This initial
operator can be based on stacking velocities or an initial macro model. In the latter
case the initial focusing operator is calculated by positioning a dipole at the focus
point followed by a forward modeling algorithm to calculate the response at the
surface. Measuring the response at the detector positions yields an operator for
focusing in detection and measuring the response at the source positions yields an
operator for focusing in emission.
The first focusing step, where the important Fresnel stack takes place, is calculated
by a convolution along the time axis of the traces in the shot record and the focusing
operator. The result of the first focusing step can be compared with the focusing
operator, indicated in figure C.1 by the move-out analysis block, and the focusing
operator and/or macro model can be updated. In the comparison the operator must
be time coincident with the focus point response in the CFP gather (principle of equal
traveltime). Combining focus point responses for the same lateral position into one
gather gives the CFP image gather which is calculated in the second focusing step.
Every CFP image trace can be compared with a ’reference trace’ to calculate the
270
C.1 Processing flow
shot
records
operator
traveltimes
scalar
image
first
focusing
step
move-out
analysis
operator
updating
second
focusing
step
coherency
analysis
image
quality
vector
image
Figure C.1 Schematic representation of the CFP algorithm. Note that the processing is done in a
shot gather stream, which makes the in and output straightforward. The ⋄ in the scheme
means that the result for more than one shot record is needed to proceed in the scheme.
coherency in the CFP image gather. This coherency analysis is based on equation
(5.3.3). The traces in a CFP image gather are added together to construct the
one-way time image (scalar image). The ⋄ in the scheme means that the result for
more than one shot record is needed to proceed in the scheme. This means in the
algorithm that either these traces must be stored in memory or on disk. For the
direct calculation of the scalar image the only traces to be kept in memory are the
traces representing the final result; the one-way time image. For the calculation
of the vector image all the traces in a CFP gather are used. The vector image is
obtained by a 2-dimensional convolution between the focusing operator and the CFP
gather.
Appendix C: Algorithms
C.2
271
Time- and Frequency-domain processing
The integral to be calculated for the first focusing step in detection is given by (after
equation (4.19))
Z
−,s
G+,∗ (x, xr )P −,s (xr , xs )d2 xr ,
(C.1)
P (x, xs ) =
∂D1
where ∂D1 is the surface the detectors are positioned on and P −,s (xr , xs ) the observed upgoing wavefield after pre-processing, to remove the direct wave and the
multiple reflections related to the surface. The Green’s function occuring in equation (C.1) can be calculated in several ways. For this thesis two calculation methods are implemented; one carried out in the time-domain and another one in the
frequency-domain.
The frequency domain method is the most accurate one because it uses equation
(C.1) directly. The only assumptions are made in the calculation of G+,∗ (x, xr ),
which is mainly determined by the accuracy of the macro model. A good method
to calculate the Green’s function is the recursive extrapolation technique based on
the extrapolation operators discussed in appendix A.
For the time-domain method the Green’s function in equation (C.1) is replaced with
its ray-asymptotic form given by
G+,∗ (x, xr ) ≃ A(x, xr ) exp (jωτ (x, xr )),
(C.2)
where τ satisfies the Eikonal equation and A the transport equation. The traveltimes
τ can be calculated by making use of a ray-tracing program, or an Eikonal solver.
√
,
The amplitudes of the Green’s function, A can be approximated by cos (φ) √jω
r
resulting in the expression
√
X
jω
−,s
(C.3)
cos (φ) √ P −,s (xr , xs )exp(jωτ (x, xr )),
P (x, xs ) ≈∆x
r
∂D1
√
where jω represents
p the 2-dimensional propagation factor, cos (φ) the directivity of
√
the dipole, r = (z0 − zr )2 + (x0 − xr )2 the distance from the focus point (x0 , z0 )
to the receiver position at (xr , zr ) and ∆x the distance between the receivers. Note
that this linear distance does not describe the length of the ray-path correctly in
inhomogeneous media.
For the time-domain calculation of one CFP trace for focusing in detection the traces
in the shot record are shifted upward with the times given by the focusing operator.
Which is equivalent to a convolution in the time domain with a time shifted delta
function. The amplitude of this delta function is defined by the amplitudes belonging
to the operator times. If the time given by the focusing operator does not fit on the
discrete time-grid of the shot record a linear interpolation to the operator time is
272
C.2 Time- and Frequency-domain processing
shot records
FFT
focusing
operators
multiplication
+
summation
interpolation
+
summation
FFT-1
first focusing step
CFP trace
Figure C.2 Difference between time and frequency processing for the first focusing step. Note that
the great computation difference is the FFT.
carried out. Optional in the first focusing step is an additional time correction with
the time defined by the traveltime between the focus point to the source position.
This additional correction performs the move-out correction on the CFP trace and
is useful in the move-out analysis.
The frequency-domain method is easier to implement, but takes much more computation time as shown in figure C.2. The shot record and the synthesis operator are
both transformed to the frequency domain and multiplied with each other. Which
is equivalent to a time convolution of the traces in the shot record and the traces
in the focusing operator. After the multiplication the summation over all traces in
the shot record is carried out and the resulting CFP trace is transformed back to
the time-domain. In a homogeneous medium the main difference between the time
and frequency method is that the time-domain method uses linear interpolation and
the frequency-domain method uses sinc functions for the interpolation to calculate
the CFP corrected traces of the shot record. For non-homogeneous media the timedomain method can only take one arrival time into account, the frequency domain
method can take in principle all arrival times into account. For the extraction of
AVO information the frequency-domain method is recommended, because it allows
the best treatment of the amplitude information present in the data.
Appendix C: Algorithms
N shots
273
D operators
first
focusing
step
NxD CFP traces
D CFP gathers
Figure C.3 Scheme of the data flow used to calculate CFP gathers. For an efficient implementation
the focusing operators are kept in memory. Every shot record gives one trace of a shot
gather.
C.3
Numerical implementation (time-domain)
The time-domain implementation of the CFP processing scheme works with any
initial focusing operator defined by traveltimes and amplitudes as given in equation
(C.2). The initial focusing operator used in the examples of this thesis are obtained
with a finite difference calculation method based on the method and software of
Vidale (1988). Vidale’s method calculates traveltimes through any velocity structure
274
C.3 Numerical implementation (time-domain)
on a 2- or 3-dimensional numerical grid, amplitudes are not calculated in Vidale’s
scheme. The amplitudes belonging to the traveltimes are calculated by determining
the shortest distance between the focus point and the receiver position, where ray
bending is not included. The traveltimes and amplitudes of the focusing operators
are stored as single traces in an operator file on disk. In the CFP calculation program
the focusing operators are read into memory and stored in an operator table.
An alternative for using Vidale’s operators in the first and second focusing step, are
focusing operators based on a root mean-square velocity, as obtained after conventional NMO analysis. The advantage is that only a root mean-square velocity model
has to be read in and the focusing operators can be calculated easily within the
program. The big disadvantage is that a homogeneous medium assumption is used
implicitly in this approach.
For every input shot record one CFP trace is calculated for all the defined focusing
operators, which results in a collection of CFP traces as shown in figure C.3. A
shot record contributes to a CFP gather if the lateral distance between the focus
point and the source position does not exceed a certain distance set by the user.
After all shot records have been treated the traces which belong to one focus point
are collected and the CFP gather is available. In the scheme it is possible to write
the already calculated CFP traces to disk and do the sorting into CFP gathers
afterwards. If not that many CFP traces are calculated it is more convenient to
keep them in memory. Optional in the processing scheme are a move-out correction
to calculate the CFP move-out panel and a mute factor to remove the ’non-causal’
events in the CFP gather, which occur before the times of the focusing operator.
For focusing in emission the traces in every common receiver gather are used to
construct the CFP trace. This means that to complete the summation along the
source coordinate of the common receiver gather all the shot records which contain
the receiver of interest have to be read in.
The second focusing step is carried out by combining time selections of the CFP
traces generated in the first focusing step for focus points defined at the same lateral
position as shown in figure C.4. The first focusing step is adjusted to this selection by
calculation only those time samples in the CFP trace which are needed in the second
focusing step. To carry out the second focusing step a synthesis operator is needed
at every time sample. The focusing operators for times (or depths) in between two
focusing operators of the first focusing step are obtained by linear interpolation of
the operator times. The interpolated operators are used to correct the events in
the CFP gather. If the correct operators are used all events in the CFP image
gather should after the time correction align at the one-way image time. Due to
the convergence of the focusing operators at higher offset a stretching is introduced
in the second focusing step. The user can set a stretch mute factor to control the
amount of stretching. The user can also control which shot record contributes to the
image gather by restricting the maximum lateral distance between the shot record
Appendix C: Algorithms
N shots
275
D operators
L positions
first+second
focusing
step
LxN image traces
L image gathers
CFP image
Figure C.4 Scheme of the data flow used to calculate CFP image gathers. Note that it is possible to
implement the first and second focusing step in an efficient way.
and the focus point.
The CFP image is calculated by a summation along the traces in CFP image gather.
The result is positioned at the lateral position of the focusing operator at the one-way
image time, or depth, of the synthesis operator. Looking at the shape of the focusing
beams as presented chapter 5, the first focusing process may allow a course focus
point sampling (∆T ) and still be accurate for the second focusing step. Typically
∆T ranges from 40 to 100 [ms].
276
C.3 Numerical implementation (time-domain)
Bibliography
Abramowitz, M. and Stegun, I. A. (1968). Handbook of Mathematical Functions.
Dover publications Inc., New York.
Alá’i, R. (1997). Improving predrilling views by pseudo seismic borehole data. Ph.D.
thesis, Delft University of Technology. in progress.
Alford, R. M., Kelly, K. R., and Boore, D. M. (1974). Accuracy of finite-difference
modeling of the acoustic wave equation. Geophysics, 39, 834–842.
Algazi, V. R., Suk, M., and Rim, C. S. (1986). Design of almost minimax FIR
filters in one and two dimensions by WLS techniques. IEEE Trans. Circuits and
Systems, CAS-33, 590–596.
Aminzadeh, F., Burkhard, N., Kunz, T., Nicoletis, L., and Rocca, F. (1995). 3-D
modeling project; 3rd report. The Leading Edge, 15, 125–128.
Berkhout, A. J. (1982). Imaging of acoustic energy by wave field extrapolation (2nd
edition). Elsevier, Amsterdam.
Berkhout, A. J. (1984). Seismic resolution: resolving power of acoustic echo techniques. Geophysical Press Ltd.
Berkhout, A. J. (1987). Applied seismic wave theory. Elsevier, Amsterdam.
Berkhout, A. J. (1992). Areal shot record technology. Journal of Seismic Exploration, 1(2), 251–264.
Berkhout, A. J. (1996a). Pushing the limits of seismic imaging: part I. Geophysics,
61, in print.
Berkhout, A. J. (1996b). Pushing the limits of seismic imaging: part II. Geophysics,
61, in print.
278
Bibliography
Berkhout, A. J. and Rietveld, W. E. A. (1995). Prestack migration in terms of
double dynamic focusing. In 65th Annual Internat. Mtg., Soc. Expl. Geophys.,
Expanded Abstracts, pages 1228–1231. Soc. Expl. Geophys.
Berkhout, A. J. and Thorbecke, J. W. (1996). Regularization of prestack data. In
DOLPHIN: Optimum Data Acquisition Technology for Maximum Image Quality,
volume 1, chapter 4. Delft University of Technology.
Berkhout, A. J. and van Wulfften Palthe, D. W. (1979). Migration in terms of
spatial deconvolution. Geophysical Prospecting, 27, 261–291.
Berkhout, A. J. and Verschuur, D. J. (1996a). CFP migration and multiple scattering. In DELPHI: From seismic measurements to rock and pore parameters,
volume 7, chapter 11. Delft University of Technology.
Berkhout, A. J. and Verschuur, D. J. (1996b). Pushing the limits of seismic imaging part III: CFP migration and multiple removal. In DELPHI: From seismic
measurements to rock and pore parameters, volume 7, chapter 10. Delft University
of Technology.
Berkhout, A. J. and Wapenaar, C. P. A. (1989). One-way versions of the Kirchhoff
integral. Geophysics, 54, 460–467.
Berkhout, A. J., Wapenaar, C. P. A., and Verschuur, D. J. (1996). CFP migration
and mode-converted data. In DELPHI: From seismic measurements to rock and
pore parameters, volume 7, chapter 12. Delft University of Technology.
Biondi, B. and Palacharla, G. (1994). 3D depth migration using rotated McClellan
filters. In Expanded Abstracts of Papers, page B45. EAEG.
Black, J. L. and Brzostowski, M. A. (1994). Systematics of time-migration errors.
Geophysics, 59(9), 1419–1434.
Blacquière, G. (1989). 3D wave field extrapolation in seismic depth migration.
Ph.D. thesis, Delft University of Technology.
Blacquière, G. (1991). Optimized McClellan transformation filters applied in onepass 3-D depth migration. In Expanded Abstracts, pages 1126–1129. SEG.
Blacquière, G., Debeye, H. W. J., Wapenaar, C. P. A., and Berkhout, A. J. (1989).
3D table-driven migration. Geophysical Prospecting, 37(8), 925–958.
Bleistein, N. (1984). Mathematical methods for wave phenomena. Academic Press,
Inc., Orlando.
Bojarski, N. N. (1983). Generalized reaction principles and reciprocity theorems
for the wave equations, and the relationship between the time-advanced and timeretarded fields. Journal of the Acoustical Society of America, 74, 281–285.
Bibliography
279
Born, M. and Wolf, E. (1970). Principle of optics. Pergamon press, New York,
fourth edition.
Bracewell, R. N. (1986). The Fourier transform and its applications, second edition.
McGraw-Hill, Singapore.
Claerbout, J. F. (1971). Towards a unified theory of reflector mapping. Geophysics,
36, 467–481.
Claerbout, J. F. (1976). Fundamentals of geophysical data processing. McGraw-Hill.
Clayton, R. W. and Stolt, R. H. (1981). A Born-WKBJ inversion method for
acoustic reflection data. Geophysics, 46, 1559–1567.
Corones, J. (1975). Bremmer series that correct parabolic approximations. Journal
of Mathematical Analysis and Applications, 50, 361–372.
Cox, H. L. H. (1991). Estimation of macro velocity models by wave field extrapolation. Ph.D. thesis, Delft University of Technology.
de Bruin, C. (1992). Linear AVO inversion by pre-stack depth migration. Ph.D.
thesis, Delft University of Technology.
de Hoop, A. T. (1988). Time-domain reciprocity theorems for acoustic wave fields in
fluids with relaxation. Journal of the Acoustical Society of America, 84, 1877–1882.
de Hoop, M. V. (1992). Directional decomposition of transient acoustic wave fields.
Ph.D. thesis, Delft University of Technology.
Devaney, A. J. (1982). A filtered backpropagation algorithm for diffraction tomography. Ultrasonic Imaging, 4, 336–350.
Esmersoy, C. and Oristaglio, M. L. (1988). Reverse-time wave-field extrapolation,
imaging and inversion. Geophysics, 1, 97–106.
Fishman, L., McCoy, J. J., and Wales, S. C. (1987). Factorization and path integration of the Helmholtz equation: Numerical algorithms. Journal of the Acoustical
Society of America, 81, 1355–1376.
Fokkema, J. T. and van den Berg, P. M. (1993). Seismic applications of acoustic
reciprocity. Elsevier, Amsterdam.
French, W. S. (1974). Two-dimensional and three-dimensional migration of modelexperiment reflection profiles. Geophysics, 39, 265–277.
French, W. S. (1975). Computer migration of oblique seismic reflection profiles.
Geophysics, 40, 961–980.
280
Bibliography
Gaiser, J. E. (1994). Optimum transformation and filter structure for explicit
finite-difference 3D migration. In Expanded Abstracts of Papers, page B42. EAEG.
Gardner, G. H. F. (1985). Migration of seismic data. In Geophysics reprint series,
volume 4. Society of exploration geophysicists.
Gazdag, J. (1978). Wave equation migration with the phase-shift method. Geophysics, 43, 1342–1351.
Geoltrain, S. and Brac, J. (1993). Can we image complex structures with firstarrival traveltime? Geophysics, 58, 564–575.
Gray, S. H. and May, W. P. (1994). Kirchhoff migration using eikonal equation
traveltimes. Geophysics, 59, 810–817.
Grimbergen, J. L. T., Wapenaar, C. P. A., and Dessing, F. J. (1996). Modal expansion of one-way operators in laterally varying media. Geophysics, page (submitted).
Hagedoorn, J. G. (1954). A process of seismic reflection interpretation. Geophysical
Prospecting, 2, 85–127.
Hale, D. (1991a). 3-D depth migration via McClellan transformations. Geophysics,
56, 1778–1785.
Hale, D. (1991b). Stable explicit extrapolation of seismic wavefields. Geophysics,
56, 1770–1777.
Hatton, L., Larner, K., and Gibson, B. S. (1990). Migration of seismic data from
inhomogeneous media. Geophysics, 46, 751–767.
Hazra, S. N. and Reddy, M. S. (1986). Design of circularly symmetric low-pass
two-dimensional FIR digital filters using transformation. IEEE Trans. Circuits
and Systems, CAS-33, 1022–1026.
Hegge, R. F. and Fokkema, J. T. (1996). Macro model updating by global inversion
of CFP operators. In 66th Annual Internat. Mtg., Soc. Expl. Geophys., Expanded
Abstracts, page accepted. Soc. Expl. Geophys.
Herrmann, P. (1992). Decomposition of multi-component measurements into P and
S waves. Ph.D. thesis, Delft University of Technology.
Hoff, J. D. (1995). 3D one-way wave field extrapolation by using 1D convolution
operators. Master’s thesis, Delft University of Technology.
Holberg, O. (1988). Towards optimum one-way wave propagation. Geophysical
Prospecting, 36, 99–114.
Bibliography
281
Hondius, C. J. (1993). Design of stress source arrays for predefined P- or Swavefields. Master’s thesis, Delft University of Technology.
Hubral, P. (1977). Time migration-some ray theoretical aspects. Geophysical Prospecting, 25, 738–745.
Hubral, P., Schleicher, J., and Tygel, M. (1996). A unified approach to 3-D seismic
reflection imaging, Part I: Basic concepts. Geophysics, 61, 742–758.
Hulshoff, A. (1993). Design of source arrays for pure P- or S-wave radiation.
Master’s thesis, Delft University of Technology.
Johnson, L. W. and Riess, R. D. (1977). Numerical Analysis. Addison-Wesley.
Kabir, M. M. N. and Verschuur, D. J. (1993). Parallel computing of the parabolic
Radon transform - Application for CMP-based preprocessing. In 63rd Annual Internat. Mtg., Soc. Expl. Geophys., Expanded Abstracts, volume 93, pages 193–196.
Kabir, M. N. N. (1997). Macro model estimation of the complex subsurface using
the common focus point technology. Ph.D. thesis, Delft University of Technology.
in progress.
Kabir, M. N. N. and Verschuur, D. J. (1996). Macro model estimation using primary
and multiple reflections. In DELPHI: From seismic measurements to rock and pore
parameters, volume 7, chapter 7. Delft University of Technology.
Kao, J., G.Li, and C.Yang (1994). Optimized 2D filters for 3D wavefield extrapolation on the massively parallel processors. In Expanded Abstracts of Papers, page
B48. EAEG.
Kato, H. and Matsumoto, G. (1982). Design of circularly symmetric 2-D FIR digital
filters using Fourier reconstruction technique. IEEE Trans. Acoustics, Speech and
Signal Processing, ASSP-30, 505–508.
Koek, E. A. (1997). Physical modeling of irregularly sampled data. Ph.D. thesis,
Delft University of Technology. in progress.
Koek, E. A., Faber, G., and Ongkiehong, L. (1996). Physical modelling of irregularly sampled data. In 58th Mtg. Eur. Assoc. Expl Geophys., Extended Abstracts,
page Session:B048. Eur. Assoc. Expl. Geophys.
Kogbetliantz, E. G. (1960). Generation of Elementary Functions, volume 1 of
Mathematical Methods for Digital Computers, pages 7–35. New York: John Wiley
& Sons.
Krebs, J. R., Fara, D. R., and Berlin, A. E. (1996). Accurate migration using offsetcheckshot surveys. In 58th Mtg. Eur. Assoc. Expl Geophys., Extended Abstracts,
page Session:X020. Eur. Assoc. Expl. Geophys.
282
Bibliography
Larner, K. and Hatton, L. (1990). Wave equation migration: two approaches. First
Break, 8(12), 433–448.
Lindsey, J. P. (1989). The Fresnel zone and its interpretive significance. The Leading
Edge, 8, 33–39.
Loewenthal, D., Roberson, R., and Sherwood, J. (1976). The wave equation applied
to migration. Geophysical Prospecting, 24, 380–399.
McClellan, J. H. (1973). The design of two-dimensional digital filters by transformations. In Princeton Conf. on Inform. Sci. and Sys., pages 247–251. Princeton.
McClellan, J. H. and Chan, D. S. K. (1977). A 2-D FIR filter structure derived from
the Chebychev recursion. IEEE Trans. Circuits and Systems, CAS-24, 372–378.
McCLellan, J. H. and Parks, T. W. (1972). Equiripple approximation of fan filters.
Geophysics, 37, 573–583.
McClellan, J. H., Parks, T. W., and Rabiner, L. R. (1973). A computer program
for designing optimum FIR linear phase digital filters. IEEE Trans. Audio Electroacoust., AU-21, 506–526.
McMechan, G. A. (1983). Migration by extrapolation of time dependent boundary
values. Geophysical Prospecting, 31, 413–420.
Mersereau, R. M., Mecklenbräuker, W. F. G., and Quatieri, T. F. (1976). Mcclellan transformations for two-dimensional digital filtering: I-design. IEEE Trans.
Circuits and Systems, CAS-23, 405–413.
Morton, S. A. and Thorbecke, J. W. (1996). Automating prestack migration analysis using common focus point gathers. In 66th Annual Internat. Mtg., Soc. Expl.
Geophys., Expanded Abstracts, pages 1262–1265. Soc. Expl. Geophys.
Mufti, I. R., Pita, J. A., and Huntley, R. W. (1996). Finite-difference depth migration of exploration-scale 3-D seismic data. Geophysics, 61, 776–794.
Nautiyal, A., Gray, S. H., Whitmore, N. D., and Garing, J. D. (1993). Stability
versus accuracy for an explicit wavefield extrapolation operator. Geophysics, 58,
277–283.
Neidell, N. S. and Taner, T. M. (1971). Semblance and other coherency measures
for multichannel data. Geophysics, 36, 482–497.
O’Brien, M. J. and Gray, S. H. (1996). Can we image beneath salt? The Leading
Edge, 15, 17–22.
Oristaglio, M. L. (1989). An inverse scattering formula that uses all the data.
Inverse Problems, 5, 1097–1105.
Bibliography
283
Parkes, G. and Hatton, L. (1987). Towards a systematic understanding of the
effects of velocity model errors on depth and time migration of seismic data. First
Break, 5(4), 121–133.
Parks, T. W. and Burrus, C. S. (1987). Digital Filter Design. New York: John
Wiley & Sons.
Pessoa, F. (1990). Het boek der rusteloosheid door Bernardo Soares. Arbeiderspers,
Amsterdam.
Powell, M. J. D. (1981). Approximations theory and methods. Cambridge University
Press.
Press, W. H., Teukolsky, S. A., Vetterling, W. T., and Flannery, B. P. (1992).
Numerical recipes in C, The art of scientific computing. Cambridge University
Press, second edition.
Rabiner, L. R. and Gold, B. (1975). Theory and application of Digital Signal
Processing. Prentice Hall.
Ralston, A. (1967). Rational Chebyshev approximation, volume 2 of Mathematical
Methods for Digital Computers, pages 264–285. New York: John Wiley & Sons.
Rayleigh, J. W. (1894). Theory of sound, volume II. Macmillan, London, 1945;
second, edition.
Rice, J. R. and Usow, K. H. (1968). The Lawson algorithm and extensions. Math.
Comp., 22, 118–127.
Rieber, F. (1936). A new reflection system with controlled directional sensitivity.
Geophysics, 1, 97–106.
Rietveld, W. E. A. (1995). Controlled illumination in prestack seismic migration.
Ph.D. thesis, Delft University of Technology.
Rietveld, W. E. A. and Thorbecke, J. W. (1994). Areal shot record technology:
macro model estimation. In DELPHI: From seismic measurements to rock and pore
parameters, volume 5, chapter 10. Delft University of Technology.
Schneider, W. A. (1978). Integral formulation for migration in two-dimensions and
three-dimensions. Geophysics, 43(1), 49–76.
Schultz, P. S. and Sherwood, J. W. C. (1980). Depth migration before stack.
Geophysics, 45, 376–393.
Sollid, A. and Arntsen, B. (1994). Cost-effective 3D one-pass depth migration.
Geophysical Prospecting, 42, 755–776.
284
Bibliography
Soubaras, R. (1996). Explicit 3-D migration using equiripple polynomial expansion
and Laplacian synthesis. Geophysics, 61, 1386–1393.
Stolt, R. H. (1978). Migration by Fourier transform. Geophysics, 43(1), 23–48.
Tarantola, A. (1984a). Inversion of seismic reflection data in the acoustic approximation. Geophysics, 49, 1259–1266.
Tarantola, A. (1984b). Linearized inversion of seismic reflection data. Geophysical
Prospecting, 32, 998–1015.
Thorbecke, J. W. (1992). Weathered layer corrections. In DELPHI: From seismic
measurements to rock and pore parameters, volume 3, chapter 4. Delft University
of Technology.
Thorbecke, J. W. (1994). Estimation of near surface effects; an alternative approach
to statics. In DELPHI: From seismic measurements to rock and pore parameters,
volume 5, chapter 3. Delft University of Technology.
Thorbecke, J. W. (1995). 3D Wave field extrapolation; a comparison between
different methods. http://wwwak.tn.tudelft.nl/ jan/articles/3Doperators.ps.Z.
Thorbecke, J. W. and Berkhout, A. J. (1993). Estimation of near surface effects.
In DELPHI: From seismic measurements to rock and pore parameters, volume 4,
chapter 3. Delft University of Technology.
Thorbecke, J. W. and Berkhout, A. J. (1994). 3-D recursive extrapolation operators: An overview. In 64th Annual Internat. Mtg., Soc. Expl. Geophys., Expanded
Abstracts, pages 1262–1265. Soc. Expl. Geophys.
Thorbecke, J. W. and Berkhout, A. J. (1995). 3-D recursive extrapolation operators
- A comparison. In 57th Mtg. Eur. Assoc. Expl Geophys., Extended Abstracts, page
Session:P094. Eur. Assoc. Expl. Geophys.
Thorbecke, J. W. and Rietveld, W. E. A. (1994). Optimum extrapolation operators
- a comparison. In 56th Mtg. Eur. Assoc. Expl Geophys., Extended Abstracts, page
Session:P105. Eur. Assoc. Expl. Geophys.
Thorbecke, J. W., Berkhout, A. J., and Wapenaar, C. P. A. (1992). Weathered
layer corrections in combination with wavefield decomposition. In 54th Mtg. Eur.
Assoc. Expl Geophys., Abstracts, pages 558–559. Eur. Assoc. Expl. Geophys.
Timmerman, J. I. H. (1993). Removal of finite aperture artefacts. Master’s thesis,
Delft University of Technology.
Bibliography
285
Tits, A. L. and Zhou, L. H. (1996). User’s guide for CFSQP version 2.4: A C-code
for solving (large scale) constrained nonlinear (minimax) optimization problems,
generating iterates satisfying all inequality constraints. personal notes from the
author.
Ursin, B. (1983). Review of elastic and electromagnetic wave propagation in horizontally layered media. Geophysics, 48, 1063–1081.
Verschuur, D. J. (1991). Surface related multiple elimination, an inversion approach. Ph.D. thesis, Delft University of Technology.
Verschuur, D. J., Berkhout, A. J., and Wapenaar, C. P. A. (1992). Adaptive
surface-related multiple elimination. Geophysics, 57(9), 1166–1177.
Versteeg, R. and Grau, G. (1991). The Marmousi experience . In EAEG workshop
on practical aspects of seismic data inversion. Eur. Assoc. Expl. Geophys.
Vidale, J. (1988). Finite-difference calculation of travel times. Bulletin of the
Seismological Society of America, 78(6), 2062–2076.
Wapenaar, C. P. A. (1991). On finite aperture artefacts. Technical report, Delft
University of Technology.
Wapenaar, C. P. A. (1996a). One-way representations of seismic data. Geophysical
Journal International, 127, 178–188.
Wapenaar, C. P. A. (1996b). Reciprocity theorems for two-way and one-way wave
vectors: a comparison. Journal of the Acoustical Society of America, 100, (December issue).
Wapenaar, C. P. A. (1996c). Short note: Reciprocity of one-way propagators.
Geophysics, ?, submitted for publication.
Wapenaar, C. P. A. (1997). Three-dimensional migration of cross-spread data:
resolution and amplitude aspects. Geophysics, 62, (accepted).
Wapenaar, C. P. A. and Berkhout, A. J. (1989). Elastic wave field extrapolation.
Elsevier, Amsterdam.
Wapenaar, C. P. A. and Grimbergen, J. L. T. (1996). Reciprocity theorems for
one-way wave fields. Geophysical Journal International, 127, 169–177.
Wapenaar, C. P. A., Peels, G. L., Budejicky, V., and Berkhout, A. J. (1989). Inverse
extrapolation of primary seismic waves. Geophysics, 54(7), 853–863.
286
Bibliography
Summary
Common Focus Point Technology
The subject of this thesis is a new method for imaging of seismic data and is based on
the systems oriented WRW model. This new imaging procedure can be considered as
a conventional Kirchhoff migration broken up into two separate focusing steps. By
dividing the imaging step into two separate focusing steps the understanding of the
seismic migration process has been made easier. The result after one focusing step,
called the Common Focus Point gather, is a very convenient domain for analysis
on seismic data. For example, traveltime and amplitude analysis can be carried
out in the CFP domain in a straightforward manner. One of the most important
contributions of the CFP technology is the possibility to estimate the correct focusing
operators without estimating a macro model. In this way a high quality seismic
image can be generated without knowing the underlying macro model.
To construct the CFP-gather from the shot records an initial focusing operator is
needed. This initial operator can be based on stacking velocities or an initial macro
model. In the latter case the initial focusing operator is calculated by positioning a
source at the focus point followed by a forward modeling algorithm to calculate the
response at the surface. Measuring the response at the detector positions yields the
time-reversed operator for focusing in detection and measuring the response at the
source positions yields the time-reversed operator for focusing in emission. Using this
initial focusing operator, a trace of the Common Focus Point gather is constructed
by a frequency-domain multiplication between the traces of the focusing operators
and the traces in the shot record followed by an integration over all the receivers
present in the shot record. The most important contribution in the integration result
is determined by the Fresnel zone which contribution is defined by the ray-path from
the source position via the focus point to the surface.
Once the complete CFP gather is built the focus point response can be compared
with the focusing operator. If the correct focusing operator is used then the principle
of equal traveltime holds, which means that the traveltime of the time-reversed
focusing operator is equal to the traveltime of the focus point response. If the
wrong operator is used the principle of equal traveltime does not hold anymore and
288
Summary
the focusing operator must be updated for a correct focusing result. The updated
focusing operator can be found in between the traveltimes of the focus point response
and the erroneous focusing operator. A convenient choice is the traveltime in the
middle of the focus point response and the focusing operator. With this updated
operator a new CFP gather can be constructed and the principle of equal traveltime
can be used again to check the result. In most cases one or two iterations are
sufficient to arrive at a correct focusing operator. Once the correct operator has
been found the second focusing step can be carried out.
In the second focusing step the information in the focus point response is added
together to construct the image. For this step the same focusing operator as used
in the first focusing step can be used again. The imaging result is positioned on the
same lateral position as the defined focus point at one-way image time or depth.
The one-way image time is defined as the time a wave needs to travel from the focus
point to the surface, where the lateral position at the surface is the same as the
focus point. Due to the tube-like shape of the focusing beam only a limited number
of focusing operators is needed in the first focusing step to illuminate the complete
subsurface.
The examples in this thesis demonstrate the new opportunities of the CFP technology in operator updating and structural imaging. By dividing the imaging process
into two focusing steps the problems in structural imaging are investigated again,
but now with the tools made available by the double focusing procedure. Based
of the CFP technology new ideas are developed for AVO analysis, reflector dip
estimation, regularization of irregular data, resolution analysis and 3-dimensional
imaging. By computing the so-called focusing beams, the influence of the data acquisition geometry on the migration result can be clearly demonstrated. Based on
the operator updating process an automatic operator updating scheme is currently
developed, which makes it possible to compute an image without user interaction.
The collection of estimated focusing operators that are generated during the structural imaging process will be used in the future to estimate the macro model by a
one-way tomographic inversion process (global inversion).
Samenvatting
Een techniek gebaseerd op metingen met een gemeenschappelijk brandpunt
In de seismiek worden technieken gebruikt om seismische metingen aan het aardoppervlak om te zetten in een afbeelding van de ondergrond. In dit proefschrift
wordt een nieuwe seismische afbeeldingstechniek geintroduceerd die gebaseerd is op
het systeem-georienteerde WRW model. De nieuwe techniek bestaat uit twee aparte
focusseringsstappen die na elkaar worden uitgevoerd. Dit in tegenstelling tot het
conventionele Kirchhoff migratie proces, dat de twee focusserings stappen tegelijkertijd uitvoerd. Het opsplitsen in twee focusseringsstappen verhoogt het inzicht in het
afbeeldingsproces. Het resultaat na de eerste focusserings stap is een dataset met een
gemeenschappelijk brandpunt (afgekort CFP) en kan gebruikt worden voor het analyseren van reflectie-tijden en reflectie-amplituden. Het is mogelijk om de correcte
focusseringsoperator te schatten zonder gebruik te maken van een voortplantingssnelheidmodel. Hiermee is het mogelijk om een afbeelding van de ondergrond te
maken zonder dat er enige kennis is vereist over de voortplantingssnelheid van die
ondergrond.
Een CFP meting wordt berekend door een focusseringsoperator toe te passen op
de seismische metingen. De focusseringsoperator is gebaseerd op een gemiddelde
voortplantingssnelheid als functie van de verticale looptijd of op een snelheidsmodel
gegeven als functie van de diepte. In dit laatste geval wordt de operator berekend door een bron op het brandpunt te plaatsen gevolgd door toepassing van een
algoritme om het golfveld, afkomstig van het brandpunt, aan het aardoppervlak
te berekenen. Een operator voor focusering in detectie wordt verkregen door het
golfveld te meten op de posities van de ontvangers, een meting op de bronposities defineert een operator voor focusering in zenden. Een CFP dataset wordt
geconstrueerd door een frequentie-domein vermenigvuldiging (in tijd een convolutie) tussen de traces van de seismische meting en de focusserings operator, gevolgd
door een optelling over de ontvangers in één meting. De belangrijkste bijdrage in
het resultaat na optelling wordt bepaald door de Fresnel-zone. De Fresnel-zone is
gedefineerd rond het golfpad dat loopt vanaf de bronpositie via het brandpunt naar
het aardoppervlak.
290
Samenvatting
Als de complete CFP dataset is opgebouwd dan kan de responsie van het brandpunt
vergeleken worden met de focusserings operator. Als de juiste operator is gebruikt
dan is de looptijd van de, in tijd gespiegelde, operator gelijk aan de looptijd van
de responsie van het brandpunt. Als er een verkeerde operator is gebruikt dan is
dit niet het geval en moet de focussering operator worden aangepast om een correct
focusseringsresultaat te verkrijgen. De correcte focusseringsoperator kan gevonden
worden tussen de looptijden van de (verkeerde) focusseringsoperator en de responsie
van het brandpunt. Een handige keuze voor de gecorrigeerde operator is een looptijd
precies in het midden van de respons en de operator. Met deze gecorrigeerde operator
kan een nieuwe CFP dataset worden berekend en het resultaat kan weer worden
geanalyseerd. In de meeste gevallen zijn een of twee iteraties genoeg om tot de
correcte operator te komen. Als de correcte operator is gevonden dan kan de tweede
focusseringsstap worden uitgevoerd.
In de tweede focusseringsstap wordt alle informatie van het brandpunt bij elkaar
opgeteld om zo een afbeelding van het focus punt te verkrijgen. Voor deze tweede
focusseringsstap moet dezelfde operator als in de eerste focusseringsstap gebruikt
worden. Het resultaat wordt op dezelfde laterale positie neergezet als het brandpunt, in een-weg afbeeldingstijd, of direct in diepte. De een-weg afbeeldingstijd
is gedefineerd als de looptijd die een golf nodig heeft om van het brandpunt naar
dezelfde laterale positie aan het oppervlak te komen. Rond het brandpunt is de
belichting niet beperkt tot het brandpunt alleen. Dit maakt het mogelijk om in
de eerste focusseringsstap een beperkt aantal brandpunten te gebruiken om toch de
gehele ondergrond voldoende te belichten. In de tweede focusserings stap worden
geinterpoleerde operatoren gebruikt om op alle diepten de afbeelding te kunnen
berekenen.
De voorbeelden in dit proefschrift laten de mogelijkheden zien van de CFP technologie voor het schatten van de operatoren en, vervolgens, het afbeelden van structuren
in de ondergrond. Door het afbeeldingsproces te verdelen in twee aparte focusseringsstappen is er een fundamenteel nieuwe manier onstaan om het afbeeldingsproces
te verbeteren. Zo zijn er nieuwe ideeen ontwikkeld voor amplitude analyse, schatting van de helling van een reflector, het regulariseren van onregelamtig gesampelde
data, het uitvoeren van een resolutieanalyse en het afbeelden in een 3-dimensionale
ruimte. De focusserende bundel geeft informatie over de invloed van de acquisitie
op het afbeeldingsresultaat. Op het ogenblik wordt er in de industrie een automatische operatorcorrectie algoritme ontworpen dat gebaseerd is op het operator
correctieschema. Dit project maakt het mogelijk om een afbeelding van de ondergrond te maken zonder interactie van de gebruiker. De verzameling van operatoren,
die tijdens het afbeeldingsproces worden verkregen, kan in de toekomst gebruikt
worden om een snelheidsmodel te schatten door gebruik te maken van een een-weg
tomographisch inversie proces.
Curriculum vitae
PERSONAL
Name:
Born:
Nationality:
Jan Willem Thorbecke
November 9 1965, Oostzaan
Dutch
EDUCATION
1978 - 1979:
1979 - 1985:
1985 - 1991:
1991 - 1991:
1991 - 1996:
’Zaanlands Lyceum’, Zaandam
’R.S.G. Broklede’, Breukelen
Applied Earth Sciences, Delft University of Technology
MSc thesis, Technical Geophysics,
”Scattering by a strip in a homogeneous medium”
Philosophy, Leiden University
Applied Physics, Delft University of Technology
PhD thesis, Laboratory of Seismics and Acoustics,
”Common Focus Point Technology”
WORK
1996 - present:
research geophysicist at Cray Research /
Silicon Graphics, De Meern, The Netherlands
292
Curriculum vitae
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement