Topographic Mapping from High Resolution Space Sensor Report

Topographic Mapping from High Resolution Space Sensor Report
Topographic Mapping from High Resolution Space Sensor
Report by Bulent Cetinkaya, Mustafa Erdogan, Oktay Aksu, Mustafa Onder
General Command of Mapping, Ankara TURKEY
Work package 4 – ‘Automatic Land Use Classification’
1 Introduction
In order to evaluate the IKONOS multispectral image for classification purposes, a series of
qualified collateral data is needed. In this study, both supervised and unsupervised
classifications are performed to produce a land cover map for evaluating purposes as the only
collateral data that we have is the land map of the area which gives limited but up to date
information of the area.
2 Unsupervised Classification
The ISODATA algorithm is used to perform an unsupervised classification. ISODATA stands
for “Iterative Self-Organizing Data Analysis Technique”. The ISODATA Clustering method
uses the minimum spectral distance formula to form clusters. It begins with either arbitrary
cluster means or means of an existing signature set, and each time the clustering repeats, the
means of these clusters are shifted. The new cluster means are used for the next iteration.
For Unsupervised Classification, six classes are chosen. The results for an urban and a rural
area are shown in images 1b and 2b. Labeling the classes was quite simple due to a well
classified image and the small number of classes.
As seen on image 1b, the infrastructure of the city, buildings and roads are well classified as
class 2 (purple). Some large flat buildings and infrastructures with very bright spectral
reflectance are assigned to a different class. Their roofs might be made using metal or a
different kind of material than the other small buildings and infrastructures. Because of the
limitations of the collateral data we cannot be sure about the true nature of these materials.
They are represented in the image by the colour tan.
Vegetation is represented by three classes due to the infrared band being sensitive to the
vegetation. Dark green, bright green and black colours are assigned to these classes. Grass is
represented by bright green, while forests are represented by dark green and black. The forest
subject to shadow and the water are usually represented by black. Bare ground with no grass
is assigned to class 4, represented in Brown.
As we concentrated on the Signature Mean Plot of the classes shown on figure 1, class 1
(black) and class 3 (dark green) have almost the same reflectance in all bands except infrared.
They both represent forest and trees. Class 3 may represent trees in shadow. Bare ground and
land with grass also have similar reflectance in almost all bands but with a little shift.
Figure 1: Signature Mean Plot of the Unsupervised Classification
3 Supervised Classification
The decision rules for the supervised classification process are multi-level:
- non-parametric
- parametric.
And there are three different parametric decision rules: maximum likelihood, Mahalanobis
distance, and minimum distance. For the supervised classification in this study, only the
maximum likelihood decision rule was used.
The maximum likelihood decision rule is based on the probability that a pixel belongs to a
particular class. The basic equation assumes that these probabilities are equal for all classes,
and that the input bands have normal distributions.
At the start, six classes were identified for supervised classification: forest, water, roads and
buildings, green agricultural land, cultivated agricultural land, and bare land.
The result was satisfactory for all classes except the water class. Some lakes and water
features on the multispectral image are mistakenly assigned to the roads and buildings class.
Another water class, labelled as lake water, was added and the multispectral image was
reprocessed for supervised classification. This gave a better result, but there are some water
features that are still not identified. The process was rerun after adding another water class,
labelled “fisher’s pond”. As a result the total number of classes is increased to eight but there
are still only six true classes, as three classes combine to give the water class. Results of the
supervised classification are shown in images 1c and 2c.
It is interesting to note that some water features are still misclassified as roads and buildings.
These are shown with red boxes on images 2a and 2c. The depth, quality, dirtiness of the
water itself, the type of the material that lies under the water and the marsh, reeds and other
vegetation in the water might give rise to different spectral reflectance. In order to investigate
this further, more collateral data would be needed.
In the supervised classification the forest is only represented by one class, coloured dark
green. Green agricultural land, cultivated agricultural land and bare land are assigned to
different classes, represented by bright green, brown, and tan respectively. Roads and
buildings are assigned to a single class, coloured purple. The result was quite satisfactory, but
the image still contains some speckle, which can be removed by filtering.
For filtering, a neighbourhood function is used. Neighbourhood functions are specialized
filtering functions that are designed for use on thematic layers. Each pixel is analyzed with the
pixels in its neighbourhood. The number and location of the pixels in the neighbourhood are
determined by the size and shape of the filter, which is user-defined.
Each filtering function results in the centre pixel value being replaced by the result of the
filtering function. The filtering function used in this case is “majority”, in which the centre
pixel is replaced by the most common data value in the neighbourhood.
At the start, a 3x3 window is selected as the kernel size for filtering. The speckle is removed
from the output classified image and the resulting image seems sharper than the original
classified image; especially when focusing on lines and buildings (image 1d and 2d). The
forest area seems much more homogeneous, but differentiation of the separate buildings is not
so clear.
The use of a kernel size of 5x5 for filtering (see images 1e and 2e) resulted in the loss of
much of the information in the classified image. However, if a smooth output if required from
the classification, this filter may be useful. The process results in a lower degree of precision
at the individual pixel scale, but may be of interest at a more general scale. For example it
could be used in forested areas to determine the general area of forest (ignoring small patches
of bare land and tracks and paths within the forest).
4 Conclusion
The multispectral Ikonos imagery proved to be capable of differentiating between six classes
of land cover, using a cartographic map as “ground truth” data. Further work would be
required to assess the accuracy of the classification with respect to a more detailed ground
truth dataset.
Figure 2: Signature Mean Plot of the Unsupervised Classification
Image 1a : IKONOS Multispectral Satellite Image (4 metre)
Image 1b : Unsupervised Classified Image (six classes)
Image 1c : Supervised Classified Image (eight classes)
Image 1d : Supervised Classified Image after neighbourhood filtering (3x3)
Image 1e : Supervised Classified Image after neighbourhood filtering (5x5)
Image 1f : Landmap (Ordnance Survey 10K raster) of the area of interest
Image 2a : IKONOS Multispectral Satellite Image (4 metre)
Image 2b : Unsupervised Classified Image (six classes)
Image 2c : Supervised Classified Image (eight classes)
Image 2d : Supervised Classified Image after neighbourhood filtering (3x3)
Image 2e : Supervised Classified Image after neighbourhood filtering (5x5)
Image 2f : Landmap (Ordnance Survey 10K raster) of the area of interest
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