The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008 
1094 
Rank 
Method 
RMSE [m] 
Visual 
evaluation 
Time [min] 
1 
SGM 
3,74 
21 
10,40 
2 
DLW 
4,53 
20 
4,48 
3 
Standard 
4,66 
19 
24,56 
4 
GraphCut 
4,74 
20 
154,22 
Table 1. Ranking of the analysed methods for generation of a 
digital surface model 
All following investigations were carried out with the resulting 
digital surface model of modified versions of the two dynamic 
programming algorithms - digital line warping and semi 
global matching which becomes necessary due to the non- 
epipolar geometry of the image pairs. So the programs have to 
use directly the orbit and attitude information provided by the 
corrected RPCs to avoid an intermediate resampling step to 
epipolar geometry. Since the ground resolution of the satellites 
is in the range of one meter the generated surface model in the 
same resolution is rather rugged in comparison to surface 
models from airborne camera or lidar data. 
Occluded pixels for which no height can be determined will be 
filled with the lowest neighbour value for visualization 
purposes of the DSM and marked as undefined for further 
processing if seen in none of the two stereo images. 
Figure 5. Digital surface model calculated for a section of 
600 m x 400 m from the Munich scene using the “dynamic 
line warping” approach 
3.3 Extracting the digital terrain model (DTM) 
Using the DSM a digital terrain model describing the ground 
can be derived. This is accomplished by calculating a 
morphological erosion with a filter size of the maximum of the 
smallest diameter of all buildings. This results in a height 
image with every pixel representing the minimum height in 
this area around the pixel. This approach already described in 
(Weidner and Forstner, 1995) fails in cases of DSMs 
containing outliers below the real terrain. Such values will 
dominate the resulting DTM. So in our processing chain the 
morphological erosion was replaced by a median filter 
returning a rather low order value. After filtering an averaging 
using the same filter size is applied to obtain a smoother DTM. 
In the Munich scene shown above the DTM simply reduces to 
a flat plane on street level. A more sophisticated example 
using the Athens scene is shown in Figure 6. 
Figure 6. Sections 1000 mxlOOO m from the Athens scene, 
left: DSM, right extracted DTM 
3.4 Creating a normalized digital elevation model (nDEM) 
Subtracting the DTM from the DSM gives a so called 
normalized digital elevation model consisting of the height of 
objects above the ground. In the Munich example the nDEM 
looks quite identical to the DSM due to the fact that the DTM 
is nearly flat in the shown area. In more hilly urban areas like 
the section of the Athens scene shown in Figure 6 the 
subsequent usage of an nDEM instead of the DSM becomes 
more important. The relation between DSM, DTM and nDEM 
is visualized in Figure 7. 
Figure 7. Profile across DSM, derived DTM and calculated 
nDEM for a section from the Athens scene (profile from the 
hill in the upper center to the center of the image; Gray-Values 
are arbitrary height units (parallaxes)) 
3.5 Creating true orthophotos 
Thanks to the rather dense DSM, the RPCs from the original 
imagery and the pansharpened multi-spectral stereo images it 
is possible to derive true orthophotos. In the extracted DSM 
pixels occluded in both stereo images were marked as