Petra Zimmermann 
  
2.3 Elements of the system 
The system derives in each step very general information or features, then there is an instance - a very small agent-like 
unit, that knows the "application and requirements", so is able to give parameters, rules and the order of combining the 
modules. As a result attributed or new features with information about quality of results are stored: That means a system 
similar to a "unit construction set" with connectors able to derive general information and featured, able to exchange 
information with other modules of the same level or from others. Our system provides a basic set of Java classes to 
import and export elevation data, images and vector data results, for visualisation and interaction. The core is easy to 
expand for new features or for new algorithms and consists in classes for image and elevation data processing and 
feature extraction (edge detectors, watershed segmentation, colour segmentation, morphological filters etc.) and classes 
to store and process the derived features. What we call features are Points2D/3D, Lines2D/3D, Polygons2D/3D etc. 
with information about used cues, algorithms, parameters and quality of the feature. By combining features derived 
from different views or cues we get additional attributes. This data structure enables us to combine 2D and 3D 
information in each step, to process top-down and bottom-up, and to integrate a knowledge base or models in each step. 
For detailed building extraction we may use the same algorithms, same data structure, and same features as for coarse 
building detection by choosing different models, rules and knowledge. Control and exchange while processing is 
realised through a set of classes able to connect between the basic classes, an operator can replace single classes. 
3. SEGMENTATION METHODS FOR AERIAL IMAGES AND DSMS 
One goal of AMOBE II is the automatic building extraction including coarse 3D description from coloured aerial 
images and DSM or DTM data. The system tries to derive as many as necessary useful ,low level* information with 
according accuracy and to fuse these information (area, pixel or edge based). We implemented different segmentation 
methods, edge-based, region based and hybrid methods for DSM data and aerial imagery. Here we present one method 
to DSM segmentation and one method to aerial colour image segmentation. 
3.1 Elevation - Blob extraction 
We use the DTM and DSM data to derive general features as aspect and slope of the terrain (Skidmore 1989). A surface 
can be approximated using a bivariate quadratic function: z 2 ax^  by^ * cxy dx ey f (1) 
  
The change of elevation in x and y direction can be = 
used to identify the direction and magnitude of the : = 
steepest gradient. These two parameters can be ve 
found by taking the partial first order derivatives of 
(1) above, with respect to x and y. The slope can be 
computed by combining the two component partial 
derivatives: 
de 
dxy 
This can be transformed and written as 
slope = arctan(Vd”+e*), (3) 
aspect = arctan(^) : (4) 
  
  
  
  
  
  
The DSM is segmented by extracting regions with 
high slope and aspect and if available checked also 
by subtracting the DTM data. So we get coarse 
"blobs", regions that give the location of objects 
that have higher elevation than the surrounding 
terrain. These objects may be buildings, bridges or 
trees etc. (Figure 5). The blobs are stored with their 
boundary polygon information, the slope and 
aspect information inside and average slope and 
aspect along the boundary (Figure 6,7). The 
accuracy of the position and shape of the blob is 
Example blob, 
see below 
  
  
  
heuristically set to a lower value than in the other 
algorithms. Figure 5: Extracted blobs through DSM information 
  
  
1066 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 
3.2 
direc 
(Fig 
brea 
conv 
For 
bour 
"Bas 
can | 
  
Figu 
ima; 
The 
info 
elev 
33 
Lon 
ima; 
edge 
filte 
leve 
chaı 
1, € 
the : 
Hor 
hon