Evaluation of SPOT 4 Image Data for Forest Cover

Evaluation of SPOT 4 Image Data for Forest Cover
Evaluation of SPOT 4 Image Data for Forest Cover Mapping
Suzanne Furby, Xiaoliang Wu and Jared O’Connell
CSIRO Mathematical and Information Sciences
CMIS Technical Report 06/155
September 2006
Executive Summary
SPOT 4 image products were obtained for three test regions to evaluate the issues that arise if this imagery
were to be used in place of the Landsat 5 imagery in the Land Cover Change Program (forest cover
mapping) of the Australian Greenhouse Office. All aspects of the forest cover mapping were considered;
including scene selection, ortho-rectification, calibration, mosaicing and thresholding to produce forest
cover maps.
The recommendations for the Land Cover Change Program are:
•
The scene selection criteria should be extended to include a low incidence angle as part of the
requirements. However, with the imagery available for the test areas and in their surrounds, the
only cloud free image often had high incidence angles. In southern Australia, complete coverage
for the 2006 epoch will be unlikely. Variable pointing angles can shift some images quite
extensively so that there are gaps between adjacent images. Selection criteria will need to include
verification of overlap with all adjacent scenes.
•
The registration of SPOT images to the rectification base is poor in mountainous areas, unless
DEMS having higher resolution than the merged AUSLIG 3 and 9 second DEMs are used in the
standard processing sequence. Localised shifts of up to 100m have been observed in the NSW test
area. Even with higher resolution DEMs, displacements of about 25m remain in some regions.
•
If SPOT imagery is going to be used operationally for several epochs, the choice of pixel
resolution relative to the existing time series of data needs to be investigated.
•
Additional Master GCP features are required as the existing features used with Landsat images do
not provide a sufficiently dense spatial coverage for the much smaller SPOT scenes. The orthorectification effort will be increase by a ‘per image’ amount corresponding to increasing the
number of Master GCPs, as well as by the increased number of images to be processed.
Approximately six SPOT scenes are required per Landsat TM scene for complete coverage.
•
Calibration of the SPOT imagery requires extensive further study. New BRDF models and
coefficients are required. As the intensity range of the SPOT imagery is much greater than for the
Landsat TM imagery, a new calibration base will need to be investigated, particularly if the data is
to be later used to monitor trends. As with the Master GCPs, the invariant target locations used in
the calibration process are too sparsely located for the smaller SPOT images. Once appropriate
methodologies have been established, the operational processing effort should only increase by the
‘per image’ amount necessary for the selection of new invariant targets.
•
In the thresholding stage of the processing, new indices will have to be derived for many
stratification zones as not all Landsat TM spectral bands are present in the SPOT image data – this
may take up to two weeks effort per map sheet. The analyses should be performed by a very
experienced team. Ideally indices are derived by considering two or more image dates to ensure
they are robust through time rather than tailored to particular conditions in a single image. All
indices derived for the first epoch using SPOT data should be reviewed when a second epoch is
available. Once indices have been established, the overall thresholding effort is increased
compared to Landsat TM data by the number of extra scenes required for complete coverage.
•
The forest cover extent and change products are not as consistent with those obtained from
Landsat 5 as the products from the Landsat 7 SLC-off data. In the NSW test area, the mapped
forest extent is smaller and the rates of change correspondingly larger.
2
Evaluation of SPOT 4 Image Data for Forest Cover Mapping
Suzanne Furby, Xiaoliang Wu and Jared O’Connell
CSIRO Mathematical and Information Sciences
1. Introduction
SPOT 4 image products have been acquired for three test regions to evaluate the issues that arise
if this imagery were to be used in place of the Landsat 5 imagery for the next update of the Land
Cover Change Program (forest cover mapping) by the Australian Greenhouse Office. Images
from the SPOT 4 satellite have been obtained in archive mode since approximately December
2005 in response to the problems encountered with the Landsat 5 system around that time.
Acquisitions from the most recent SPOT satellite – SPOT 5 – must be pre-ordered, making SPOT
4 a more practical alternative to Landsat TM data, at least for coverage during the 2006 epoch.
All aspects of the forest cover mapping are considered including scene selection, orthorectification, calibration, mosaicing and thresholding to produce forest cover maps. The
suitability of the image data for other products such as sparse cover mapping, plantation type and
canopy density will be considered in a follow-up study reported separately. Unless indicated
otherwise in the text, all processing was performed according to the standard methodology for the
Land Cover Change Program as described in Furby (2006).
Other sources of image data are also being evaluated. These include Landsat 7 SLC-off, IRS and
CBERS imagery. The evaluations will be reported separately and a summary comparison report
produced.
2. The Test Areas
The test areas are in New South Wales, Tasmania and Western Australia as shown in figure 1.
Figure 1: Approximate location of the test areas
3
The New South Wales test area, shown in figure 2, was selected to include a black soil
stratification zone so that forest / non-forest discrimination can be evaluated in one of the most
challenging environments. It also includes a region with significant terrain effects to allow
evaluation of registration, BRDF and terrain illumination correction issues in an ‘extreme’
environment. The selected area is in the centre of map sheet SH56. The Landsat WRS path/rows
are 89/81 and 90/81. The SPOT path/rows are 387/410, 388/410, 389/410, 390/410.
Figure 2: SPOT 4 images of the New South Wales test area in map sheet SH56, bands 2,4,3 in BGR.
The yellow lines show the stratification zone boundaries used in the Land Cover Change Program.
The white lines show the boundaries of the four SPOT 4 scenes. The large black ‘gaps’ are regions
where cloud has been masked.
The Tasmanian test area, shown in figure 3, was selected to include mountainous areas as well as
an agricultural environment where discrimination between cropped paddocks and plantations can
be difficult. The selected area is just to the east of the centre of map sheet SK55. The Landsat
WRS path/rows are 90/89 and 91/89. The SPOT path/rows are 384/434, 385/434, 386/434 and
387/434.
Figure 3: The Tasmanian test area in map sheet SK55. The yellow lines show the stratification zone
boundaries used in the Land Cover Change Program. The white lines are the boundaries of the four
SPOT 4 scenes.
4
The Western Australian test area, shown in figure 4, was chosen to include a significant region of
new plantations as well as some of the wheat belt tree cover that is close to the 20% canopy cover
cut-off used in the forest definition. The test area is relatively flat. The selected area is in the
north-east of map sheet SI50. The Landsat WRS path/rows are 111/83 and 112/83. The SPOT 4
path rows are 317/ 416, 318/416, 319/416 and 319/415.
Figure 4: The Western Australian test area in map sheet SI50. The yellow lines show the
stratification zone boundaries used in the Land Cover Change Program. The white lines are the
boundaries of the four SPOT 4 scenes.
3. Scene Selection Issues
Unlike the Landsat series of satellites, the SPOT satellites can point to the left and right of nadir.
This provides more opportunity to acquire cloud-free imagery by allowing acquisition of a region
from adjacent orbital paths. SPOT 4 overpasses are being collected and archived. The archive
images have a variety of pointing angles.
The standard scene selection criteria were expanded to include an incidence angle (pointing angle
of satellite) close to zero (nadir) to minimise occlusion and between scene BRDF effects. Often, it
was necessary to sacrifice incidence angle to obtain an image with reasonably low cloud cover.
The effects are expected to be largest between images from adjacent paths rather than from
images down a single path. Hence, as far as possible, test images were selected as an east-west
row of four scenes.
The image archive was searched for several path/rows surrounding those eventually selected for
the test areas to obtain the best possible sequences of test data. For the broader areas searched,
although there appeared to be many images in the archive, most had very large areas of cloud
cover and/or significant missing data. Generally there was at most one suitable image, not a
choice of dates. In southern Australia, complete coverage for the 2006 epoch will be unlikely.
5
Another issue with scene selection was to ensure overlap (or at least no gap) between adjacent
scenes. Variable pointing angles can shift some images quite extensively so that there are gaps
between adjacent images. Selection criteria will need to include verification of overlap with all
adjacent scenes.
Finding suitable imagery for Western Australia proved the most difficult. Path 320, in a swath
down the centre of the state, appears to be entirely missing from the archive. The compromise
was to use three horizontal scenes, 317-319/416, and one scene vertical to these, 319/415.
Table 1: Selected SPOT 4 Images for the Western Australian Test Area
Path/Row
Date
Cloud
Inc. Angle
317/416
26/01/2006
Partly
-13.8
318/416
05/02/2006
Clear
+3.2
319/416
21/01/2006
Clear
-13.8
319/415
10/02/2006
Clear
+13.4
Scene selection for Tasmania was relatively easy. A large back archive is available for the
region. As is typically the case with this area, many of these scenes contained cloud so incidence
angles were not ideal.
Table 2: Selected SPOT 4 Images for the Tasmanian Test Area
Path/Row
Date
Cloud
Inc. Angle
384/434
06/01/2006
Mostly clear
+7.7
385/434
06/01/2006
Some cloud
+10.7
386/434
17/02/2006
Some cloud
-26.5
387/434
06/01/2006
clear
-13.1
There is a relatively large back archive of data for NSW also, but only a few images from 2006.
Hence sub-optimal green images were selected for two scenes. Again, due to cloud cover, scenes
with large incidence angles were chosen.
Table 3: Selected SPOT 4 Images for the New South Wales Test Area
Path/Row
Date
Cloud
Inc. Angle
387/410
22/12/05
hazy
-1.2
388/410
10/03/06
clear
-30.4
389/410
01/02/06
Mostly clear
+25.8
390/410
22/12/05
Most clear
-24.4
4. Raw Image Quality Issues
The raw images appear noisier than path level Landsat TM images. The pushbroom sensor
system means that each pixel within a line is recorded by a different sensor. Figure 5 shows a
sample from one of the Western Australian images over an area of dense forest. The image
appears to have some geometric patterns. The variation in intensity values within the forest is
6
generally small compared to the differences between forest and non-forest cover. There should be
little overall effect on forest / not forest discrimination, but there may be some effects at the edge
of forest blocks and in areas with forest density around the 20% canopy cover cut-off. Although
not being formally tested at this time, there may be issues in using texture calculated from these
images for sparse woody cover monitoring. It should be noted, however, that this noise effect is
not obvious visually in the ortho-rectified image data.
Figure 5: Bands 2, 3 and 4 (BGR) from 317/416 26 January 2006. This is an area of relatively dense
forest. A regular pattern can be seen. The area shown is approximately 3km square.
Noise stripes such as that in the centre of figure 5 appear in several of the images obtained. Such
areas can be readily masked, although masking is a manual process and if the noise stripes are
7
common it will increase effort required ‘per image’. Several of the test images contain such
stripes.
The order of the image bands in the supplied imagery is XS3, XS2, XS1, SWIR (i,e 3, 2, 1, 4).
Although the image bands can be automatically reordered, it would be simpler to request that the
data be supplied in band order for operational work.
5. Ortho-rectification Issues
The images were initially ortho-rectified using the standard procedures applied to the Landsat
data used in the Land Cover Change Program, i.e. using the Year 2000 base, merged AUSLIG 3
and 9 second DEMs and the Master GCPs files used for the 2005 Update.
Landsat TM band 7 is typically used in the correlation matching calculations to locate Master
GCP features in the overlap images. As an equivalent spectral band is not available in the SPOT
imagery (see figure 8), matching SPOT band 4 to Landsat TM band 5 and SPOT band 2 to
Landsat TM band 3 were both evaluated. Equivalent results both in numbers of GCPs matched
and GCP locations were obtained.
An output pixel resolution of twenty five metres was used to match the Landsat imagery to which
the SPOT imagery will be compared. If SPOT imagery is going to be used operationally for
several epochs, the choice of pixel resolution relative to the existing time series of data needs to
be investigated.
The issues that arose are:
• The Master GCPs files did not contain enough features to provide sufficient GCPs to
register individual SPOT images. It was necessary to add GCPs manually for all test
images.
• Large regions of cloud in the 2000 rectification base, for example in Tasmania, in
combination with a cloudy SPOT image, can cause insufficient common clear area for
GCP selection. An alternative base image was used for some of the Tasmanian images.
• In mountainous areas, particularly with large satellite incidence angles, the AUSLIG
DEM is too coarse. Significant areas with displacements of up to 100m (4 pixels) were
found.
• Even with higher resolution DEMs, very localised displacements remain in some of the
images that appear to be terrain related.
Twenty seven GCPs spread uniformly across the image area are considered sufficient to
adequately register the Landsat images (approximately 185km square images). The Master
features files typically deliver between sixty and one hundred well matched GCPs for each
(single) Landsat TM image. However, the SPOT images cover only about one sixth of a Landsat
TM image. Only ten to fifteen well matched GCPs were obtained over the test images (for all
combinations of image bands tested). Comparisons showed that better registration to the base
was obtained by manually adding GCPs so that a minimum of twenty five to thirty GCPs were
used in the model fit. Operationally, during the first epoch for which SPOT imagery is used the
Master GCP and Master Check GCP files will need to be revised for most areas, requiring a small
‘per image’ increase in effort for the ortho-rectification processing.
8
A further issue with GCP selection arose in Tasmania. There are a couple of regions in this map
sheet where the 2000 rectification base has substantial cloud cover. These cloud patches tend to
cover only 15-20% of a TM (single scene) image but, depending on location, can be a substantial
proportion of a particular SPOT scene. Restricting GCPs to only the cloud-free area in the 2000
based produced poorly registered SPOT images. An image from another epoch that was cloudfree in the region overlapping the SPOT scene was used. This problem would be restricted to
parts of Tasmania and far north Queensland only. It would be necessary to nominate a second,
alternative base to be used in such regions.
The biggest issue encountered during the ortho-rectification process was poor registration in areas
of terrain, particularly for images with large incidence angles. Figure 6 shows one such example
from the NSW set of images. Even with extra GCPs there were errors up to 100m (4 pixels) in
the registration in some valleys that could not be removed. Using the NSW state DEM in place
of the usual AUSLIG DEM produced an image with improved registration. Many problem
regions were resolved completely using the higher-resolution DEM and the remaining shifts were
limited to about one pixel. Smaller shifts were observed in the Tasmanian images using the
AUSLIG DEM (no more that 1-2 pixels). Again the registration was visibly improved by using
the state DEM. The area covered by the WA images is comparatively quite flat. In areas where
the registration showed shifts of 1/2 to 1 pixel, using the Land Monitor DEM produced worse
rather than better results.
The only area of significant terrain for which a higher resolution DEM is not readily available for
this study is south-east Queensland. Testing with the NSW images showed that the results using
the SRTM-DEM are almost identical to those from the state DEM. However, the SRTM-DEM
has numerous null pixels in areas of water bodies (particularly rivers and some lakes) as well as
smaller holes in some mountainous areas. It is also quite noisy. Some pre-processing would be
necessary so that it could be used operationally for ortho-rectification.
A feature of the observed misregistration caused by terrain is the affected areas tend to be quite
localised (3-4 km square regions) and can be surrounded by areas with similar terrain that are
well registered, i.e limited to areas of poor DEM quality. Although the four Tasmanian images
are well registered to the base, small localised pockets of misregistration of ½ to 1 pixel remain in
the NSW and Western Australian images after using the best available DEMs. It appears that the
ortho-rectification is more sensitive to terrain issues and that the registration of SPOT imagery to
the Landsat 7 base may not be quite as good as can be obtained for other Landsat imagery.
Localised regions of misregistration have implications for QA as well. With the Landsat images,
regions of misregistration caused by bad or too few GCPs tend to cover much larger geographic
areas (a whole corner or edge of an image) and are identified very rapidly. Identifying problem
areas in SPOT images would require much more intensive scrutiny of the processed images.
9
(a) AUSLIG DEM
(b) State DEM
Figure 6: Ortho-rectified 389/410 1 Feb 2006 image (green) overlaid on Year 2000 base (red)
using the AUSLIG and NSW state DEMs. The area shown is approximately 5km square. The
displacement between the images in the top display is 100m (4 pixels). A small displacement remains
when the state DEM is used.
10
6. Calibration Issues (including terrain-illumination correction)
The standard calibration process consists of three distinct steps:
• top-of-atmosphere and BRDF corrections;
• invariant target atmospheric check/correction; and
• terrain-illumination correction, if required.
The issues that arose were:
• The intensity values in the SPOT images have a much greater dynamic range than Landsat
data (2-3 times in the visible bands). To avoid compressing the data range, and hence
reducing the discrimination between forest and non-forest cover, floating point values
were rescaled to the full 0-255 data range for writing to image files.
• The viewing geometry of the SPOT images is such that new BRDF kernels as well as
coefficients may well be required. The current data were insufficient for testing the
validity of the current kernels and coefficients and a research activity will need to be
undertaken if SPOT imagery is used operationally. Coefficients were estimated to match
the test images to the calibration base using the current kernels.
• As with the Master GCPs, the invariant targets are distributed over the Landsat TM scene
area (and rarely uniformly). There are too few, if any, targets located in most of the SPOT
test images. New targets were selected for all of the test images; however most of the
good bright pseudo-invariant targets are saturated in the first two image bands. A common
gain and offset (usually estimated from pooled data) was used to align the data for each
test area with the (scaled) base.
Table 4 show the intensity range for one of the NSW images and the corresponding area in
calibration base. Particularly in the visible bands, the range of intensity values in the SPOT
image data is much greater than in the Landsat 7 ETM+ data. Similar differences were seen for
all images in each test area. A new calibration strategy is required to avoid compressing the
intensity range of the SPOT image data, with subsequent loss of discrimination, if it were adopted
operationally. For the tests conducted here the data were rescaled at each step in the calibration
process.
Table 4: SPOT and Landsat 7 ETM+ Intensity Ranges
Image Band
SPOT Intensity Range
Landsat 7 ETM+ Base
388/410 10/03/06
Intensity Range
SPOT 1 / TM 2
57 - 254
15 - 69
SPOT 2 / TM 3
40 - 254
10 - 91
SPOT 3 / TM4
19 - 254
15 - 151
SPOT 4 / TM 5
12 - 254
5 - 197
SPOT 4 imaging sensors use the pushbroom scanning system which is different from the rotating
mirror scanning system of the Landsat TM sensors. These sensor characteristics, together with
slight spectral range differences (see figure 8), create new calibration issues. Firstly, unlike for
Landsat sensors, the scaling parameter to convert the output from the earth-sun distance
correction for each band to reflectance is unknown for SPOT 4. Secondly, the BRDF kernels
used for Landsat data are not the optimal choices for SPOT 4 data due to the viewing geometry
differences (SPOT 4 has a wider field of view compared to Landsat). Thirdly, the BRDF kernel
11
coefficient estimation (for new or current kernels) requires broad scale SPOT 4 coverage (e.g. a
third of the continent).
For all SPOT 4 images in the three study areas, a reflectance scaling parameter for each band was
estimated that produced data in a plausible range for the BRDF calculations. With insufficient
test data to derive new BRDF kernels, the current Landsat kernels were applied and coefficients
were estimated scene-by-scene by matching directly to the Landsat calibration base image using
sites in forested areas. This approach is a short-term solution for the study areas but is not
recommended for large scale and operational processing.
At least ten, but ideally twenty to thirty invariant targets are considered adequate for the
atmospheric check / correction. The invariant target location files typically contain only twenty
to forty targets for each (single) Landsat TM image area. Although it is desirable for the targets
to be dispersed as widely as possible across the image, in practice the locations tend to be
somewhat more clumped. As the SPOT images cover only about one sixth of a Landsat TM
image, few, if any, of the targets were located over the SPOT images. New targets were selected
manually for each image. Operationally, this will create a ‘per image’ increase in effort for the
calibration processing.
The data in table 4 also show that each of the image bands saturates at 254 for the example image.
This occurred for most of the test images, with up to five percent of the image being saturated.
Bright targets that didn’t saturate were very difficult, if not impossible, to find.
Using the invariant targets to calibrate each SPOT image directly to the calibration base tended to
make one or more of the images more different from, rather than more similar to, its adjoining
images. It is unclear whether this is due to inadequate BRDF corrections or other factors and was
not investigated further. A common or pooled gain and offset correction was applied to all of the
images within a test area.
If SPOT imagery is to be used operationally, a much more detailed investigation of calibration
issues needs to be conducted using significantly more image data than considered here, of the
order of a good proportion of at least one state. However, once this research establishes the
appropriate corrections and parameters, the operational processing effort will only increase by the
‘per image’ amount necessary for the selection of new invariant targets.
7. Mosaicing Issues
No difficulties arose in the mosaicing stage of the processing. As there are more images in each
map sheet, the processing will take a little longer, but there are no additional issues to consider.
8. Thresholding Issues
So far the thresholding processing has only been applied to the NSW and WA test areas.
8.1 NSW Test Area
The NSW test area is shown again in Figure 7. Stratification zone 7 contains the black soils.
Discriminating between forest and non-forest cover is very difficult in this zone. More omission
12
and commission errors are made in stratification zones with this soil type than are typically
observed in other zones. Some of these persist through the multi-temporal processing. The other
stratification zones are relatively straightforward unless the imagery is particularly green or the
sun angle is particularly low.
5
7
6
4
6
3
Figure 7: SPOT 4 images of the New South Wales test area in map sheet SH56, bands 2,4, 3 in BGR.
The yellow lines and numbers show the stratification zone boundaries used in the Land Cover
Change Program. The white lines show the boundaries of the four SPOT 4 scenes. The large black
‘gaps’ are regions where cloud has been masked.
The Landsat TM thresholds are listed in table 5. Each set of indices includes at least one index
using Landsat TM bands 1 or 7 (red in table 5). As is shown in the comparison of SPOT 4 and
Landsat TM image bands in figure 8, there are no SPOT equivalents to these image bands. New
indices must be derived from the SPOT data.
Training sites of forest and non-forest cover were selected from image inspection using the 2005
forest extent map as ground-truth. As at least one new index had to be derived for each zone, the
TM-equivalent indices were compared to new SPOT-specific indices to determine whether better
indices could be derived. Table 6 shows the best indices for each stratification zone. Indices
displayed in green are the Landsat TM indices formed from the equivalent SPOT bands. No better
indices could be derived for zones 5 and 6 from the SPOT images. Indices that performed better
on these particular SPOT images were derived for zones 3 and 4 but the overall improvement in
forest / non-forest discrimination was small. The improvements tended to involve masking nonforest paddocks which were spectrally different from forest rather than separating spectrally
similar forest and non-forest regions.
Zone
3
4
5
6
7
Table 5: Landsat TM Indices for Each Stratification Zone
Index 1
Index 2
Index 3
B2 + B4 + B5
B3 – B5 + 3B7
B2 + B4 + B5
B3 – B5 + 3B7
B2 + B4 + B5
B4 – B7
B2 + B4 + B5
B4 – B7
B2 + 2B3 – B4 - B5 + 2B7
B3 + B4 + B7
B1 + B2 + B4 –B5 + B7
13
Zone
3
4
5
6
7
Table 6: New SPOT Indices for Each Stratification Zone
Index 1
Index 2
Index 3
4B1 – B2 + B3 + B4
6B1 – 6B2 - B3 + 2B4
4B1 – B2 + B3 + B4
6B1 – 6B2 - B3 + 2B4
B1 + B3 + B4
B3 – B2
B1 + B3 + B4
B3 – B2
B1 + B2 + B4
B1 + B3
6B1 – B3 - 2B4
* % !
+
" #
" #
" #
'
&&
&&
!
!
$% !
!
(
Figure 8:
" # )&
&'
!
!
Comparison of SPOT 4 and Landsat 7 TM image bands
Image matching to the 2005 single-date forest cover probability image provided good thresholds
for all but one sub-zone. Manual thresholds were used for this sub-zone in the black soil zone.
The overall correspondence with the base forest cover probability image was broadly as would be
expected from Landsat data. The same levels of omission and commission errors appear in the
black soil zone (7), although the particular cover types causing confusion were a little different.
If SPOT data is to be used operationally, new indices will have to be derived for most
stratification zones – this may take up to two weeks effort per map sheet. The analyses should be
performed by a very experienced team. Ideally indices are derived by considering two or more
image dates to ensure they are robust through time rather than tailored to particular conditions in
a single image. All indices derived for the first epoch using SPOT data should be reviewed
when a second epoch is available.
Once indices have been established, the overall thresholding effort is increased compared to
Landsat TM data. Stratification zones are intersected with the image date boundaries to
14
determine sub-zones. Thresholds are derived for each sub-zone separately. As there are
approximately six SPOT images for every Landsat TM image, there will be approximately six
times as many sub-zones to consider and potentially (in the worst case) up to six times the effort
required.
8.2 WA Test Area
The test area covers the eastern edge of the state forest, including some areas of new plantations
and harvesting, and a typical wheatbelt area. The thresholding in these stratification zones is
relatively straightforward unless the imagery is particularly green.
Only one index for one of the stratification zones uses Landsat TM bands 1 or 7. A new index
was obtained for this zone by comparing index displays formed from the indices from the
adjacent zones.
Image matching to the 2005 single-date forest cover probability image provided good thresholds
for all but one sub-zone. Manual thresholds were used for this sub-zone which contained a
number of new plantations. The overall correspondence with the base forest cover probability
image was broadly as would be expected from Landsat data, suggesting that deriving new indices
is not essential. One sub-zone showed slightly higher probabilities on some non-forest cover than
expected, however the multi-temporal processing corrected the probabilities.
8.3 Tasmanian Test Area
The test area covers the central northern part f the state, including areas of agriculture, plantation
forestry, wilderness and alpine vegetation. Difficulties are often encountered in separating green
agricultural paddocks from young plantations where the agricultural and forest regions meet.
Otherwise, the thresholding in these stratification zones is usually relatively straightforward
unless the imagery is particularly green.
All of the stratification zones have common indices. The first of these is ‘Band 3 + Band 5 +
Band 7’ for which no equivalent is available using SPOT 4 image bands. During the analyses to
derive indices from Landsat TM data for these stratification zones, the index ‘Band 3 + Band 5’
was found to perform almost as well, for which an equivalent is available using SPOT 4 image
bands. Acceptable results were obtained for all zones using the substitute index, without the
need to derive new indices.
Image matching to the 2005 single-date forest cover probability image for the agricultural and
forest zones and the 1998 single-date forest cover probability image for the wilderness and alpine
zones provided good thresholds for most sub-zones. Cloud cover in both the 2006 and historical
imagery and smaller sub-zones due to smaller SPOT 4 image extents meant that it was harder to
define matching windows that contained little land cover change but were still representative of
the cover over the whole sub-zone. Thresholds were derived manually for such sub-zones.
The overall correspondence with the base forest cover probability image was broadly as would be
expected from Landsat data, suggesting that deriving new indices is not essential. Similar levels
of commission errors were observed on green agricultural paddocks near forestry operations.
15
9. Comparison of Thresholding Results with Landsat 7 SLC-off and Landsat 5 Image Data
Forest cover probability images have been created for the 2006 epoch using the SPOT 4 data as
described in this report. Forest cover probabilities were also created from Landsat 5 data using
the standard methodology and from Landsat 7 SLC-off data as described in Furby (2006).
The image dates for the Landsat 5 imagery are:
Path/Row
89/81
90/81
Table 7 (a): Landsat 5 TM Image Dates for NSW
Date
Cloud
08/04/06
10% cloud
14/03/06
Mostly clear
Path/Row
111/83
112/83
Table 7(b): Landsat 5 TM Image Dates for WA
Date
Cloud
08/04/06
30% cloud
14/03/06
Mostly clear
Path/Row
90/89
91/89
Table 8(c): Landsat 5 TM Image Dates for Tas
Date
26/02/06
02/03/06
Cloud
40% cloud
70% cloud
The image dates for the Landsat 7 SLC-off imagery are:
Path/Row
89/81
89/81
90/81
90/81
Table 8 (a): Landsat 7 SLC-off Image Dates for NSW
Date
Cloud
Purpose
15/03/2006
Clear
Primary
31/03/2006
5%
Fill
07/04/2006
Clear
Primary
10/02/2006
40%
Fill
Path/Row
111/83
111/83
112/83
112/83
Table 8(b): Landsat 7 SLC-off Image Dates for WA
Date
Cloud
Purpose
21/01/2006
25%
Primary
06/02/2006
25%
Fill
17/03/2006
Clear
Primary
13/02/2006
25%
Fill
Path/Row
90/89
90/89
91/89
Table 8(c): Landsat 7 SLC-off Image Dates for Tas
Date
Cloud
Purpose
23/04/2006
55%
Primary
07/04/2006
50%
Fill
24/01/2006
40%
Primary
Each of the forest cover probability images was added to the existing sequence of probability
images from 1972 to 2005 for multi-temporal processing. Common error rates were applied for
16
the 2006 epoch as there was no evidence to suggest that the probability images from any of the
three data sources were more or less accurate than usual. Forest extent and change maps were
created from the outputs of the multi-temporal processing according to the usual procedures.
Figure 9 shows the forest cover products for a region in the black soil stratification zone in the
NSW test area (zone 7 in figure 7). The Landsat 5 TM image for 2006 is displayed in the top left.
The SPOT 4 image for 2006 is displayed in the bottom left. The three forest cover extent maps
are displayed together in the top right of figure 9. The forest extent map derived from the
Landsat 5 data is displayed in red. The forest extent map derived from the Landsat 7 data is
displayed in green and the forest extent map derived from the SPOT 4 data is displayed in blue.
Where these products coincide, the display appears white (forest) or black (non-forest). Colours
indicate that the area is labelled as forest by only one or two of the products. In particular, yellow
shows a where the products from Landsat 5 and Landsat 7 are identical, but the product from
SPOT 4 differs. Red, green and blue indicate that the area is labelled as forest in only the Landsat
5, Landsat 7 or SPOT 4 product respectively. The image in the bottom right of figure 9 is an
overlay of the 2005-2006 clearing layers from the three products. Again the Landsat 5 derived
product is in red, Landsat 7 in green and SPOT 4 in blue.
Yellow (red + green) indicates regions where the results from the SPOT 4 images are different to
those from the Landsat 5 and Landsat 7 images (which are consistent with each other). Similar
differences are observed throughout the test area. As a consequence clearing rates for 2004-2005
and 2005-2006 are higher when calculated from the SPOT 4 products. Similar trends were
observed in the WA test area.
Tables 5 to 7 show the area of forest and change for the NSW test area calculated from the three
products for the more recent epochs. The observation that the 2006 forest extent is similar for the
two Landsat products is confirmed. The SPOT 4 product shows a smaller 2006 forest extent and
both higher clearing and revegetation rates in the 2005-2006 interval. Most of the extra
revegetation in the SPOT 4 product appears along the edges of forest areas, particularly in the
more mountainous regions. Some of it is due to the registration not being quite as good when
extreme satellite pointing angles are combined with mountainous terrain.
Epoch
2000
2002
2004
2005
2006
Table 5: Forest Extent Comparison
Landsat 5
Landsat 7
Forest Extent (ha)
Forest Extent (ha)
901612.9
901554.2
903661.7
903371.4
905869.4
905343.1
912433.2
912809.1
914660.2
916070.6
SPOT 4
Forest Extent (ha)
901303.5
903129.6
904437.8
909352.4
910981.1
Table 6: Area of Vegetation Loss (Clearing Layer) Comparison
Interval
Landsat 5
Landsat 7
SPOT 4
Clearing (ha)
Clearing (ha)
Clearing (ha)
2000 - 2002
11961.8
12230.8
11922.4
2002 - 2004
9223.0
8992.2
9054.2
2004 - 2005
11597.5
12210.6
12388.1
2005 - 2006
7270.3
6633.9
9970.6
17
Figure 9: Top Left: 2006 Landsat 5 TM image, bands 3, 5, 4 in BGR. Bottom Left: 2006 SPOT 4
image, bands 2, 4, 3 in BGR. Top Right: Overlay of forest extent maps for 2006 for Landsat 5 (red),
Landsat 7(green) and SPOT 4 (blue). Bottom Right: Overlay of 2005-2006 clearing maps for
Landsat 5 (red), Landsat 7(green) and Spot 4 (blue).
Table 7: Area of Vegetation Gain (Revegetation Layer) Comparison
Interval
Landsat 5
Landsat 7
SPOT 4
Regrowth (ha)
Regrowth (ha)
Regrowth (ha)
2000 - 2002
14010.6
14048.0
13748.4
2002 - 2004
11440.6
10964.0
10362.4
2004 - 2005
18161.3
19676.7
17302.8
2005 - 2006
9497.4
9895.4
11599.3
It should be noted that, as illustrated in figure 10, all regions of land cover change observed in the
NSW test area were detected in all three products. The regions in which the SPOT and Landsat
18
products differ tend to be edge pixels, narrow features and the less dense forest cover. As none of
the high resolution Ikonos imagery overlaps this test site (and if it did it would show 2002/2003
cover rather than 2006 cover), it is only possible to say which is most consistent rather than most
accurate for the areas that differ. For change detection, consistency is an important consideration.
March 2006
April 2006
January 2005
December 2005
Figure 10:
Top Left: 2006 Landsat 7 TM image, bands 3, 5, 4 in BGR. Bottom Left: 2005 Landsat 5 TM
image, bands 3, 5, 4 in BGR.
Top Centre: 2006 Landsat 5 TM image, bands 3, 5, 4 in BGR. Bottom Centre: 2006 SPOT 4 image,
bands 2, 4, 3 in BGR.
Top Right: Overlay of forest extent maps for 2006 for Landsat 5 (red), Landsat 7(green) and SPOT
4 (blue).
Bottom Right: Overlay of 2005-2006 clearing maps for Landsat 5 (red), Landsat 7(green) and
SPOT 4 (blue).
10. Conclusions
When considering SPOT image data, the scene selection criteria should be extended to include a
low incidence angle as part of the requirements. However, with the imagery available for the test
areas and in their surrounds, the only cloud free image often had high incidence angles. Generally
there was at most one suitable image, not a choice of dates. In southern Australia, complete
coverage for the 2006 epoch will be unlikely. Variable pointing angles can shift some images
quite extensively so that there are gaps between adjacent images. Selection criteria will need to
include verification of overlap with all adjacent scenes. Approximately six SPOT scenes are
required per Landsat TM scene for complete coverage.
19
The registration of SPOT images to the rectification base is poor in mountainous areas, unless a
DEM with higher resolution DEMs than the merged AUSLIG 3 and 9 second DEM is used in the
standard processing sequence. Localised shifts of up to 100m have been observed in the NSW
test area. Even with such DEMs, displacements of about 25m remain in some regions.
Additional Master GCP features are required as the existing features used with Landsat images do
not provide a sufficiently dense spatial coverage for the much smaller SPOT scenes. An output
pixel resolution of twenty five metres was used to match the Landsat imagery to which the SPOT
imagery was compared. If SPOT imagery is going to be used operationally for several epochs,
the choice of pixel resolution relative to the existing time series of data needs to be investigated.
The ortho-rectification effort will be increase by a ‘per image’ amount corresponding to
increasing the number of Master GCPs, as well as by the increased number of images to be
processed.
Calibration of the SPOT imagery requires extensive further study. New BRDF kernels will need
to be derived as well as new model coefficients. Much more data than was obtained for the test
areas is required to establish appropriate models (the Landsat models and coefficients were
derived from about thirty scenes covering the whole of Queensland). As the intensity range of
the SPOT imagery is much greater than for the Landsat TM imagery, a new calibration base will
need to be investigated, particularly if the data is to be later used to monitor trends. As with the
Master GCPs, the invariant target locations used in the calibration process are too sparsely
located for the smaller SPOT images. Once appropriate methodologies have been established, the
operational processing effort should only increase by the ‘per image’ amount necessary for the
selection of new invariant targets.
In the thresholding stage of the processing, new indices will have to be derived for most
stratification zones as the not all Landsat TM spectral bands are present in the SPOT image data.
This may take up to two weeks of effort per map sheet. The analyses should be performed by a
very experienced team. Ideally indices are derived by considering two or more image dates to
ensure they are robust through time rather than tailored to particular conditions in a single image.
All indices derived for the first epoch using SPOT data should be reviewed when a second epoch
is available. Once indices have been established, the overall thresholding effort is increased
compared to Landsat TM data, in the worst case by the number of extra scenes required for
complete coverage. Cloud cover in both the 2006 and base imagery and smaller sub-zones due to
smaller SPOT 4 image extents can make it harder to define matching windows that contain little
land cover change but are still representative of the cover over the whole sub-zone, requiring
greater manual estimation of thresholds.
The forest cover extent and change products from SPOT 4 are not as consistent with those
obtained from Landsat 5 or Landsat 7 SLC-off data. The mapped forest extent is smaller and the
rates of change correspondingly larger.
11. References
Furby, S. L. (2006), Documentation for the 2005 Update of the Forest Cover Mapping for the
Australian Greenhouse Office Land Use Change Program, CSIRO Mathematical and Information
Sciences Technical Report 06/43.
Furby, S. L. and Wu, X. (2006), Evaluation of Landsat 7 SLC-off Image Data for Forest Cover
Mapping, CSIRO Mathematical and Information Sciences Technical Report 06/154.
20
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