Ulrich Thoennessen 
  
The presence of buildings in the scene can be verified in almost the entire “building expectation area”. Only a small 
part at the top of the largest building is missing, due to shadowing effects. However, the weighted means of the 
extracted object parts differed significantly from the truth heights taken from detailed building ground plans. The 
calculated heights tended to be about 5m below the expected values, even in segments which fitted almost perfectly the 
referring truth data. This seems to be surprisingly considering the better results mentioned above achieved for the 
extended grassy-areas inside the airfield traffic net. A possible reason for this effect is that the signal response from 
buildings suffers more from artefacts like multi-backscattering or layover. This behaviour is object of further studies. 
On the other hand, in the “non-building expectation area” a hint to a recently constructed building is found. It fits like a 
symmetric twin to an existing building in the middle on top of the scene detail. It came out that indeed the building was 
built up in the time period between the production of the paper map and the INSAR measurement campaign. 
The large region in the middle of the question-mark-shaped building contains a shadow region and was obviously 
under-segmented in the first iteration. In the refinement step this region is further analysed (Figure 6). The histogram 
shows no bimodal distribution. However, the distribution of the heights indicate the presence of two different height 
levels. In the region growing step, starting at the shadow area, two large regions can be segmented, besides some small 
regions nearby the shadow. In Figure 6e the final result is shown. The small regions are filtered out and the remaining 
two segments became filled. 
B0 aû up 120 
a) Median Filtered Heights b) Height Distribution c) Histogram of Heights 
      
   
d) Result of Region Growing e) Final Result 
Figure 6: Histogram of Heights and Region Growing Result of a Scene Detail Containing Shadow 
6 CONCLUSIONS AND FUTURE WORK 
The approach of improving the interpretation of InSAR height information, using a model-based segmentation process 
in the intensity data, achieved promising first results. The visual impression is improved compared to the noisy original 
data. 
After geocoding, the smoothed InSAR height data can be used for change detection purposes, at least for extended 
buildings. But the results indicate that the approach is limited to coarse scene descriptions only. For image based detail 
analysis, like roof reconstruction, the data is still too inaccurate. If a further improvement of the accuracy is achievable 
in the future, e.g. by better motion compensation incorporating more sophisticated INS, more detailed object models 
become appropriate. Gradient-based segmentation methods can be used for example to detect sloped objects like gable 
roofs. In the present state, the approach is restricted to flat terrain. In the case of undulated terrain, surrounding areas 
have to be examined to determine the elevation of buildings over ground. 
  
334 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part Bl. Amsterdam 2000. 
13 
Sc) 
pre 
So 
(n 
Sti 
Ste 
UL 
Wi