)vement and 
] an Inertial 
and Moreira | 
  
ith AeS-1 
n 
e processing 
Ts). In SAR 
rorithm. The 
3PS and INS 
: the harmful 
-off between 
resolution of 
ric processor 
netry. Also a 
geometry of 
This is done 
of figure 2). 
trimmed. To 
also can be 
is was made 
> Swiss Jura 
d open areas, 
graphic map 
ng the height 
“Topography 
1999), This 
smaller. The 
is sufficient, 
Andreas Keim 
  
  
Figure 2: Terrain-geocoded SAR scene of an area 
near Solothurn with superimposed topographic map 
(scale 1 : 10,000) 
(© Vermessungsamt d. Kantons Bern, Switzerland) 
4 MAP GENERATION 
For the map production as described in this paper, two information extraction methods are applied: visual interpretation 
and a supervised classification. 
To accerelate and simplify map production the first step after basic data preparation is to investigate the maximum 
number of distinguishable classes. This was realized by programming a classification software optimized for the high 
resolution SAR data. The used input data for the software are fully geocoded SAR images and DEMs. Typical classes 
of interest in topographic mapping are water, 
SAR : Area 
Image Selection 
   
   
   
   
forest, build-up and open areas. After the user 
had choosen training examples inside the 
graphical user interface (GUI) for the classes 
of interest an artificial neutral network (ANN) 
  
  
Training Set. | ^^ : S 3 
Selection : classifier is designed. The used ANN classifier 
UNO : : is the well-known Multilayer Perceptrom 
; ; (MLP) (Rumelhart et al. 1986). An interactive 
Feature Classifier A Classification Classified Il . . fi di 
Selection imus TY Cif maps component allows inspection of intermediate 
  
  
QUT | S results and enables feedback to training set 
selection and classifier learning, thus an online 
FRE learning capability is established (Fig. 3). The 
: borders of the classes of the improved final 
E sd . results are vectorized, stored in individual 
eda dir o Pel ours layers and outputted in DXF format. This 
7 vector data can be integrated e.g. in a DTP 
program or cartographic information system 
Figure 3: Block diagram of the ANN classifier (CIS) and combined with the data achieved by 
manual screen digitizing. 
  
  
  
  
  
In a further step of the InNSAR processing chain the third dimension of the terrain, represented by contour lines and hill- 
shading, must be produced and embedded into the map. To automatize the extraction of the contour lines from the 
existing geocoded InSAR DEM (see chapter 3), a software system was developed by Schmieder and Huber (2000). Due 
0 the fact, that microwaves with short wavelength’ (e.g. X-band) cannot penetrate into buildings and dense vegetation 
like forests, the contour lines represent only the surface of that areas which lead to an error in height. With complex 
algorithms its possible to extract that areas and correct the DEM there (Huber/Schmieder 1997). The output product is 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B1. Amsterdam 2000. 175