The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008 
1096 
However this object and outline extraction step needs some 
more sophisticated algorithms. All approaches shown in 
literature for the extraction of buildings and other objects from 
laser DSMs do not work well for the rather rugged DSM 
generated by stereo matching of the VHR satellite stereo 
images. Some solutions presented in earlier works (Krauß et. 
al., 2007) will not work automatically in highly urban areas 
like the Athens scene and so completely different approaches 
have to be developed. 
3.8 Object modeling 
For the simple modeling of the extracted objects following 
base models are used: 
• Model “ground” (class “low”, any type of vegetation) 
• Model “tree” (class “high” and “vegetation”) 
• Model “building” (class “high” and “no vegetation”) 
Figure 12. Simple models used 
The “ground” is inserted as a height field extracted from the 
DTM with an optional texture directly from the true 
orthophoto. 
“Trees” are described by a crown diameter and a treetop height 
extracted from the classification and the DSM respectively. 
“Buildings” are represented as prismatic models. In the future 
the prismatic models will be split to cuboids with optionally 
parametric gabled roofs. 
3.9 Representing the object models through geometric 
primitives and exporting in a suitable 3D format 
The coarse models will be represented by geometric primitives. 
A height field derived from the DTM for “ground” (one for the 
full scene, textured from the true orthophoto), an ellipsoid 
supported by a cylinder for trees and rectangular vertical walls 
following the extracted circumference and a horizontal 
polygonal roof in the first version. A texture may be extracted 
from the original images by projecting the resulting polygons 
backward using the RPCs. The optionally textured geometric 
primitives have to be exported into a suitable 3D vector format. 
The automatic export step of the chain already works well. 
Required inputs are simply the DSM, the true ortho photo 
matching the DSM, and the objects in form of 2D vector 
outlines. These outlines represent the two elevated classes: 
trees will be marked by circles, polygons represent buildings 
or parts of buildings with the same roof slope. In the Athens 
example these outlines were generated manually and only flat 
roofed buildings occurred in the scene (see Figure 13). 
In the export step automatically the height of the objects is 
extracted from the outlines and the DSM. A totally new digital 
terrain model will be created from the DSM by cutting out all 
elevated objects marked by the 2D outlines and consecutively 
interpolation and smoothing. The ground object will also be 
textured by the true ortho photo. 
Figure 13. Orthorectified section from the Athens scene 
(UTM projection) with manually marked trees (green circles) 
and building outlines (orange polygons) 
Figure 14 shows the automatically generated 3D model with 
the textured ground and elevated tree- and building-objects 
from the lower left quarter of the Athens scene in Figure 13 
using a VRML viewer. 
Figure 14. 3D view automatically generated from the DSM 
and the manually marked trees and building footprints from 
the Athens scene, size 500 m x 500 m 
4. SUMMARY AND OUTLOOK 
In this paper a first version of a processing chain for the 
automatic extraction of three-dimensional city models directly 
from high-resolution stereo satellite images is described. The 
chain elements are already implemented but some steps 
require further optimization. For example the DSM generation 
needs some major improvement. Also the methods developed 
in previous works for building extraction fail to work in some 
complex urban areas. So the main future work will focus on