International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
39 
Segmentation is a well established and more complicated 
method of extracting features. There are many algorithms 
available and a number were tested for this work. These are 
listed with comments in Table 1. All have the problem of 
defining parameters which can only be carried out manually. 
Algorithm 
Image 
Source 
Comment 
MUM 
SAR 
NA Software 
(Cook et al. 
1994) 
Initial oversegmentation 
followed by merging 
RWSEG 
SAR 
NA Software 
(White 1991) 
Detects edges and grows 
regions within them 
REGSEG 
SPOT 
Kai and 
Muller, 1991 
Region growing from 
seed points. Edges can 
be used if required. 
OPTISEG 
SPOT 
Ruskoné and 
Dowman, 
1997 
Developed from 
REGSEG and gives 
better results 
Table 1. Algorithms used for extracting polygonal features. 
The algorithms discussed above were tested on SPOT and SAR 
images of the same area which is Istres in Southern France. The 
original images are shown in Figure 3. 
Figure 4 shows the automatic thresholding applied to the images 
and the post processed images after clutter removal. We can see 
here a number of well-defined polygons, but also some areas 
which are not so well defined. Figure 5 shows the result of using 
a. SAR Image of Istres 
homogeneous patch extraction. The images with clutter 
removed are not shown, but the small patches are removed. 
The results are different from those from thresholding. In the 
case of the SAR data the results are similar but those for SPOT 
and quite different and indeed are better. 
The four segmentation algorithms were also tested on SAR and 
SPOT data and the results shown in Figure 6. In all cases many 
more polygons were extracted than from the thresholding and 
homogeneous patches method. Some post processing was 
carried out to merge similar regions and to remove small 
patches. The RWSEG algorithm tends to oversegment the 
images, thus requiring more post processing. No pre-processing 
was required on the SAR images but the SPOT images had to be 
normalised and smoothed. As with the SAR images, the effects 
of oversegmentation from using REGSEG had to be reduced by 
post processing but OPTISEG was optimised to extract larger 
features and only small patch removal was necessary. 
Some important results emerge from these tests. First, it is clear 
that only the thresholding algorithm is automatic and that it is 
the method which gives the most basic result. All the other 
algorithms require user intervention to obtain optimum results. 
It is also clear that different algorithms give different results, all 
of which may not be suitable for patch matching. Other tests, 
not shown here show that the results are scene dependent. The 
homogeneous patch extraction was the algorithm which was 
simplest to use and which worked reasonably well on all 
images. This may therefore be an acceptable compromise for a 
near automatic system. 
b. SPOT image of Istres 
Figure 3. Original SAR and SPOT images.