(2) 
of how 
ttering 
ibution 
ed as: 
(3) 
e SAR 
erence 
s to a 
Thomas Damaseaux 
  
4. The high resolution airborne interferometric SAR AES-1 
The data for this work were acquired with the high resolution airborne interferometric SAR AES-1, built and 
operated by Aero-Sensing Radarsysteme GmbH. The main system parameters of the AES-1 radar are 
summarized in table 1. 
  
  
  
  
  
  
  
  
  
X-Band P-Band 
operating frequency 9.35 - 975 GHz 380 - 450 MHz 
polarisation HH HH 
system bandwidth 400 MHz 70 MHz 
pulse repetition frequency 16 kHz 16 kHz 
ground resolution (range x azimuth) |0.5mx0.5m 2.5m x 0.5m 
flight velocity 50 -200 m/s 50 -200 m/s 
flight altitude above sea level 500 m - 12000 m 500 m - 12000 m 
Table 1. System parameters of the AES-1 radar 
5. Information Extraction 
The model of the real world gained from the SAR overflight (two antiparallel flight directions) should describe 
the reality regarding surface coverage and the height information as well as possible. The illustration below 
(Figure 6) shows the simplified process chain from an image of the real world to a classification result. 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Natural |. | Sensor 2 InSAR Post Feature Feature Decision| |Dempster 
— => = => => => — 
Pattern Flight Process. Process. Extract. Selection Maker Shafer 
Conflict Ea Context > Class. 
Fusion Filtering Result 
  
  
  
  
  
  
  
Figure 6. Process chain from a natural pattern to a classification result 
After the InSAR processing the data are prepared for the information extraction; this step is called post- 
processing. At first the X-band image is processed with a Lee filter [Lee, 1981] in order to reduce the speckle. 
A further step entails geocoding and radiometric correction of the scenes [Holecz, 1993]. Since these processes 
are dependent on a DEM, it has to be generated in the required cartographic reference system at the beginning 
of the process. So far the following products, also called primary features, are available for an image analysis: 
- SAR image - DEM - Coherence image 
The primary features, however, are not sufficient if the SAR images are to be interpreted adequately. In order to 
improve the model, additional information is derived from the primary features. This step is called feature- 
extraction [Dutra et al, 1998]. In it new features are obtained which are extracted with the help of a texture 
analysis of the SAR image(s) and the coherence. These features contribute significantly to the improvement of 
the classification result [Schistad Solberg et al, 1997]. 
Texture can be analysed in different ways: 
- Local Statistics Features 
- Co-occurrence matrices [Haralick et all, 1973] 
- Laws filter [Laws, 1980] 
All features obtained by feature extraction now form a d-dimensional feature space. If d is high, the problem 
arises that the accuracy of the classification decreases. This phenomenon is known as curse of dimensionality" 
[Bishop, 1995]. Therefore the best subset of m features from the set of d possible features must be found, where 
m « d [Dutra et al, 1998], [Huber et al, 1998]. For this the Jeffreis-Matusita distance (JMD) [Swain et al, 1978] 
can be used, which makes a statement on the statistical separability of n classes in a feature space between 
multivariate Gauss distributions. Although most of the features are not Gauss distributed, the JMD can 
nevertheless be used as a measure of the separability. 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part Bl. Amsterdam 2000. 57