Thomas Damaseaux 
  
The aim is now to derive the likelihoods 1, for each desired class and each pixel p at the position (x, y) from 
the m-feature subset. This will be done separately for exactly the same cutting of each flight direction. For this 
a multivariate Gaussian maximum likelihood classifier (MLC) [Swain et al, 1978] is used. 
Now one can fuse the likelihoods 1, with their appropriate coherence using the Dempster-Shafer Theory of 
Evidence (DSTE) [Klein, 1993] to get new likelihoods 1, and a weight of conflict k for each pixel p. 
Substantially this method is based on collecting data (likelihoods 1, , coherence) to solve a certain problem 
(land cover classification) by assigning probabilities to the data and fusing them with the DSTE to obtain a 
more exact overall probability. The coherence is used here as an expert knowledge. Coherence can be divided 
into stable land cover, such as man-made or bare soil, and unstable land cover, such as water, forest or 
agricultural areas. In this connection the coherence can support or dilute the likelihoods 1, of the different 
classes and so contribute to more exact likelihoods 1 , . 
: With a conflict-weighted method one can fuse 1, and k from each flight direction to a new set of likelihoods 
1, for each pixel p. By using two different flight directions for the classification one can so minimize the 
influences of the relief on the SAR-specific recording geometry. 
The likelihoods 1 ; can finally be improved with an algorithm including the context (Potts model) [Besag, 
1986] of a pixel p. 
In the end the above algorithm leads to the classification result which is a necessary base for a topographic 
map. 
As a next step one had to extract height information in form of contour lines from the DEM as shown in 
Figure 7. 
  
  
  
  
Natural Sensor 2 InSAR DEM DEM Con. lin. Contour 
— — — -— — 
Pattern flight dir. Process. optimat. extract. Lines 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Figure 7. Process chain from a natural pattern to contour lines 
The DEM developed with the help of the SAR interferometry is now the basis for the height information. 
Before one can extract the contour lines following aspects have to be taken into account: 
- Typical InSAR mistakes (phase-unwrapping) have to be analysed and minimised. 
- Optimum mosaiking between the two different flight directions and overlapping tracks have to be performed. 
- Since we have only information from the surface (X-Band) and not from the ground we have to estimate the 
height of trees, houses and other features. These heights have to be deduced from the DEM in order to have 
the exact height levels from the ground. 
- To smooth the DEM some filter operations have to be performed. 
Now the contour lines can be extracted. 
As Figure 8 shows the topographic map information is a combination between the classification result and the 
contour lines in a certain equidistance. Both pieces of information are the result of the information extraction 
from interferometric SAR data. 
  
  
Class. + Contour = Top. 
Result Lines Map 
  
  
  
  
  
  
  
Figure 8. Generating topographic maps 
6. Mapping the Edelsberg area 
The Edelsberg area belongs to the Bavarian alpine foothills [Hofmann, 1970] with three different 
geomorphological units: 
- the flysch zone of the Edelsberg 
- the Helvetikum to the north of the Edelsberg 
- the northern border of the Lime Alps in the south of the Edelsberg 
  
58 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B1. Amsterdam 2000.