Peter Lohmann 
  
  
“4 5729500.0 
Legend 77] 5729500.0 
139,6 m 
111,0 m 
  
8 = 
5729000.0 25m ] 5729000.0 
  
      
  
4] 5728500.0 
i i 5728500.0 
2585000.0 2585500.0 
2585000.0 2585500.0 
Figure 3: Test area with (left: color coded heights; right: filtered points using DTMCOR) 
    
  
  
  
2584500.0 
The DSM is made up of 990.375 points with a grid width of 1m. About 10.000 points in the regular grid contain no 
height value, (black points) and are missing in the original data set. A total of 186.867 points or 18,9 % have been 
filtered by DTMCOR (Figure 3 right). By comparing the color coded image (Figure 3, left) with the filtered points 
(Figure 3, right) the structure and shape of the peaks also becomes visible. However, the number of filtered vegetation 
points is less than the number of the eliminated points. Relatively large areas of non-topographic points are detected by 
the filter algorithmn. This can cause problems in interpolating the openings. Therefore the adequate selection of the 
filter setttings is very important, especially the tolerance with respect to the moving plane, /,, and /,, have a major 
influence the results. The higher the tolerance factors the less points will be eliminated. Consequently, this may cause 
that erroneous elevations resides in the final DTM. 
8.1 LINEAR PREDICTION VERSUS DUAL RANK FILTER 
A comparison between the results obtained by linear prediction and dual-rank-filtering has been carried within another 
test site. The area (see figure 4) is made up of a freeway, a high building and some forests. Large differences in 
elevation can be observed in a region close to the freeway. The size of the area is about 157.000 m?. The DSM consists 
of 156.565 points, about 1.000 points of the grid have no elevation value. Especially one part of the building shows 
black areas without any elevation data. 
The filtering of the test site with help of DTMCOR was done with the parameters mentionned above. 15.464 points 
(9,88 %) have been classified as vegetation or building points. 7715 points have been deleted by the trend removal and 
7749 points (4,95 %) exceed the tolerance against the predicted surface. Figure 4 shows the results after filtering. It can 
be noticed that the forest areas have not been filtered completely. Especially in the region south of the freeway not all 
vegetation points have been detected. The freeway with ist ramp on both sides is still preserved. The breaklines and 
edges of the terrain are not eliminated, only in the neighborhood of the bridge a few points are classified as not 
belonging to the terrain due to big changes in height. However the mid column of figure 4 shows unsatisfactory results 
in the case of the big building which is not removed totally. The reason is that the choosen grid size of 7 meters is too 
small. Therefore the algorithmn has to be modified to locally adapt the grid size to the size of the buildings. 
The results of dual rank filtering (Figure 4 right column) show very good results in case of the filtering of buildings and 
ramps. Forest areas are filtered very rigerously. The final interpolation sometimes yields a pattern like structure but in 
general this filter approach works quite well. Out of the 156.565 elevation values 13.225 (8,4%) have been detected as 
non-topographic points and are consequently removed and interpolated. 
6. CONCLUSION 
The obtained results demonstrate very promising results for forested areas using the approach with linear prediction. 
However the parameter settings have to be investigated and tested beforehand. As could be shown the filtering of big 
buildings requires the use of rather large grid sizes, which on the other hand are not well suited in areas of undulated 
terrain, because it is smoothed too much. Therefore it is difficult to work with one set of parameters in areas of different 
topography. The most sensitive parameter is the size of the grid. A possible solution to this problem is the 
implementation of the technique of a progressive sizing determining the grid size locally as a function of the terrain 
curvature. A problem specific to forest areas, is the vast filtering of points, which in consequence results in problems 
with the interpolation of these gaps. 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 545