Rob Dekker 
  
logarithmically scaled image intensities. Logarithmic scaling has the advantage that it simplifies adaptive filtering and 
detection by transforming the distribution of the speckle noise into a symmetrical distribution. Another thing is that the 
intensity difference is now measured in dBs. Change detection is sensitive to a global intensity differences, which can 
be due to a difference in the radar system gain. In case of the method of Dekker (1998) a global difference has to be 
eliminated. The OS-CFAR detector is not sensitive to this phenomena because it detects against the local background. 
After thresholding, all change pixels connected to less than a certain number of adjacent pixels are rejected, to select the 
changes that fit at least one resolution cell. 
Change detection was performed between the images of 29 January 1996 and 14 December 1998. For both algorithms 
the threshold was set to 3.719 times the standard deviation, corresponding to 9.73 dB and a false alarm rate of 0.01%. In 
case of the method of Dekker (1998) a global difference of 1.75 dB was subtracted. All change pixels connected to less 
than three adjacent pixels, were rejected. Figure 3 shows the result. The left image is a colour composite of both images 
in cyan and red, which results in unchanged objects having a grey tone and new objects being red. The right image 
shows the result of the method of Dekker (1998). The result of the OS-CFAR method is not shown. It performs well but 
the method of Dekker (1998) turned out to be more accurate. From figure 3 we see that a colour composite gives a good 
overview of the urban growth, but that the change detection algorithm gives a far more accurate picture. The result 
shows a high correlation with the ground truth: the most rapidly growing areas show most new objects. The fact that the 
incidence angle of the radar matches with the average tilt angle of the roofs, contributes to this. Another advantage of 
change detection maps is that they are very suitable as input for the estimation of the number of new houses, and the 
increase of the population index. 
  
Figure 3. Colour composite of the images of 29 January 1996 (cyan) and 14 December 1998 (red), and the detected new 
objects projected on the image of 14 December 1998 (right). The less-dense expanding areas show most new objects. 
4 SEGMENTATION AND CLASSIFICATION 
Land use classification can be split up into segmentation and classification. In the first step an image is divided in 
segments or regions with a certain uniformity. In the second a land use class is assigned to every region by evaluating 
the features extracted from the region. In another study than the one described here, segmentation of SAR data using 
textural features was examined (Dekker 2000). Segmentation was done by optimised region growing. Among the 
textural features that were studied were the standard deviation, higher order moments, grey level co-occurrence matrix 
features, grey level difference vector features, wavelet based features, fractal based features and others. The lacunarity, 
  
64 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B1. Amsterdam 2000. 
  
afr 
nun 
is g 
199 
spe 
witl 
of « 
Sliu 
inte 
Uni 
seg 
seg