International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
  
From the real SAR image three pyramid layers are calculated 
and afterwards segmented, using an adaptive threshold method 
(Levine & Shaheen, 1981). The segments are transformed to 
vector polygons. Starting from the pyramid level with the low- 
est resolution, the polygons, which represent peak areas, are 
analysed. Neighbouring polygons representing peak areas, grow 
together. Analysed polygons, with area sizes not exceeding a 
maximum area and not undershooting a minimum area, are 
selected and the search is continued in these polygons on the 
next higher resolution pyramid level. The thresholds of the 
maximum and minimum area sizes are derived from the simu- 
lated model. Those thresholds are getting more exact in pyramid 
levels of higher resolution. Using this simple method, only a 
few polygons are left as possible building positions. 
These positions are afterwards analysed using a pixel-wise com- 
parison of the simulated and the SAR image. The pixel with the 
highest concurrence is considered as the real position of the 
landmark building. For each building or object searched in the 
image, one coordinate pair can be determined. 
  
     
Figure 15. Simulated image of the original 3D-model (left) and 
of the modified 3D-model (right) 
er Sté 
   
   
   
5 ; — ol "m T : du e eus 
Figure 12. DOSAR image (left) and orthoimage (right) 
  
After the automated geo-referencing, the position of the build- 
the automated geo-referencing was not very accurate, it is nec- 
essary to determine the position of the analysed building block 
again. This is done using just the pixel-by-pixel comparison in a 
small subset. Sometimes this pixel-by-pixel method is not nec- 
essary or working, for example if huge changes occur. In this 
case a good initial geo-referencing and a DEM is needed to de- 
termine the accurate position of the building block. 
The simulated image of the 3D-model should afterwards be 
compared to the real image. Areas with high-values in the real 
  
Figure 13. DOSAR image (left) and orthoimage (right) SAR image, but low values in the simulated image may indicate 
the existence of new objects. Otherwise areas with low-values 
In Figure 12 and Figure 13 the good geo-referencing of the data — i the real SAR image and high values in the simulated image 
is visible. The area around the St. Bernhardus church was refer- may indicate missing objects. The length of the layover and 
enced using the model of the church shown in Figure 5, which shadow areas’ wre determined by the height of an object. 
is quite erroneous. The result is comparable to a manual refer- Changes in the shape of those areas can be a result of the 
encing. Even experienced operators have problems finding cor- change in the shape of the analysed objects. 
responding points using SAR and optical data. The railtracks fit 
quite well, also the buildings do fit well. The kerb shows some 
misplacement. 
The accuracy which can be achieved is difficult to determine, 
because, not many points are directly comparable. The esti- 
mated accuracy is around 1.5m in the vicinity of the church, 
which is equivalent to 6 pixel of the SAR image or 3 times the 
3-dB resolution of the SAR system. 
5. CHANGE DETECTION 
The exact geo-referencing of the different datasets is a pre- 
requisite for the change detection. The building block used as 
example for the change detection is shown in Figure 14. The 
model has been modified. The black building is not real. It is 
too high and the shape of the roof is wrong. In Figure 15, the 
difference between the SAR simulation of the original and the 
modified model of Figure 14 can be seen. DOSAR image 
476 
ing block is known. Because of topographical distortions, or if 
  
Interne 
In Figi 
block 
mentec 
the bu 
model. 
the bu 
in the « 
  
Figure 
The er 
in a k 
SAR 1 
examp 
analys 
area o 
kerb n 
and th 
the mi 
area, i 
analys 
Altoge 
ages 0 
model: 
using 
simula 
basis f 
is poss 
ful ge 
remair 
mainly 
A prer 
referer 
data a 
differe 
lookin 
geome 
datase 
ered. 
A SAI 
the SZ 
simula 
detecti 
cally ¢ 
3D-me 
the S/ 
150m 
duced 
For a : 
useful