Jochen Schiewe 
  
Looking at the spectral texture the statistical analysis significantly proves that the highest values can be found With 
buildings because their breaklines are imaged much sharper than those of trees or other objects (figure 6). Applying the 
same principle as described with the slope gradient it is possible to separate very well between all object classes, Cog. 
sequently, spectral edges could give evidence about the outline of buildings. But in practise, a couple of topological apg 
geometrical processing steps have to be applied in advance, e.g. for the reduction of the extracted regions by shadow 
areas. 
  
  
  
  
  
threshold (= u+10) 
  
  
  
  
  
  
i 
  
  
  
  
  
  
  
s SE SSugksususastnsnannsesnpany 
  
  
  
  
   
Building Wooded area 
Figure 6. Spectral texture profiles of buildings resp. wooded area. 
In order to find a final decision on the separation of buildings and wooded areas within this second stage all determined 
membership probabilities are compared with the goal to find a significant if not unique majority for one object class. 
This procedure also allows for the failure of one of the indicators (like the NDVI in our case) which is quite typical for 
real type applications and the use of different sensor configurations. It is also possible to integrate additional indicators 
(like shadow information) into the estimation. 
With that the desired detection of buildings or wooded areas is finished and the following boundary description phase 
can be performed on basis of existing approaches like that of Sahar and Krupnik (1999), Shi et.al. (1997) or Fischer 
et.al. (1998) that partly also incorporate building hypotheses. 
3.4 DSM control and generation 
In order to improve the accuracy and - even more important - the reliability of automatically derived elevation models, 
an integration of other DSMs (e.g., through the merge of multiple correlations coming from a multiple image overlap) 
or of image and value-added data (in order to check the height behaviour of underlying object classes and their neigh- 
bourhood) can be used. 
3.4.1 Discussion of previous work. Most studies are concentrating on the detection of blunders within single or 
multiple DSMs by means of geometrical parameters only. For instance, Lohmann and Koch (1999) define blunders by a 
certain deviation of a local plane. This method works well for point blunders, but worse for actual linear jumps like 
breaklines. Baltsavias et.al. (1995) present a merge of multiple correlations from a manifold image overlap on basis of 
the so-called “Figure of Merit” (FoM) which unfortunately gives unsatisfactory results. 
3.4.2 Our approach. Concluding from the results concerning DSM control presented so far we highly recommend 
the integration of image data and image-derived information in order to try to check heights according to associated 
structures or objects rather than on relying on geometrical parameters only. In the following we will examine spectrally 
homogeneous regions which are critical areas for automatic matching algorithms. In this context, Krupnik (19%) 
proposes an interpolation of all interior heights. However, from our point of view homogenous regions should be treated 
in more detail, at least by grouping them into the following three categories (figure 7): 
e Some regions (especially waters) demand not only for an interpolation but also for a constant height (i.e., Ah =0); 
e Some regions (especially shadows) do not allow any prediction about their heights — these have to be further 
processed or omitted; 
e All other classes might be modelled through a plane or might show considerable height variations — depending on 
the underlying object type. 
  
812 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 
  
In ord 
tion of 
üre ex] 
and ne 
* FR 
CO 
m 
So far 
implen