ISPRS Commission III, Vol.34, Part 3A ,Photogrammetric Computer Vision*, Graz, 2002 
stand close to buildings or buildings have been built on slopes. 
Only few buildings were not segmented at all. 
  
Correct Passable Not segmented 
Rural area 78 13 9 TR 
Urban area 85 10 5 
  
Table 3-1. Quality of the segmentation [%] of the laser scanner 
data 
  
The segmentation result is shown in Figure 3-2. To separate 
houses and trees that have been enclosed in one segment, 
sublevels were created. Sublevel means that smaller segments 
are calculated within the segments of the super level. For the 
sublevel, an aspect image of the interpolated laser scanner data 
is generated. (Equation (2)) AX and AY are the average 
elevation changes in the coordinate directions. Thus, areas with 
a high elevation change density have a strong textured 
appearance. They consist of many small segments. Whereas 
homogeneous areas ideally have only one segment. Referring to 
houses, it is assumed that there will be one segment per roof 
side. This information will also be used in the house modelling 
procedure, which is still under development. 
aspect = arctan AX (2) 
AY 
To each segment of every file in the project, eCognition 
supplies additional information, such as standard deviation, 
mean, minimal, maximal values and shape and neighbour 
information. These attributes are important parameters for the 
building detection. Hence, a slope and a laplacian filter image 
of the laser scanner data is added to the eCognition project in 
order to increase the information potential. The segmentation 
result of the height image and its sublevel (the aspect image 
based level), are converted into vector data with the attributes 
supplied by eCognition. 
In a second eCognition project the modified pixel map is 
segmented. The segments, only representing buildings and parts 
of the lettering, are also exported as vector data. The next 
chapter will refer to it as the map-segment-file. 
3.3 Results 
Although both study areas have a different structure, the 
preparation for the segmentation process was the same. In 
matter of house segments, it is interesting that commercial, light 
industry, and multi-family residential areas are segmented well. 
Only the roof segments of multi-family residential areas appear 
to be insufficient. For single-family residential units the results 
of house and roof segments in both study areas are satisfactory 
and equivalent. 
The mean and standard deviation values of the segments in the 
laplacian and slope image seem to be very useful in the 
detection analysis. 
A disadvantage of the segmentation with eCognition is still the 
time-consuming generation and export of vector data. The 
segmentation process cannot be fully automated, as for each 
scene individual parameter settings are necessary. The scale 
parameters of the laser scanner data and aspect image 
segmentation are for example higher for city areas. Another 
condition for eCognition is that the images for a given project 
have to have the same dimension. 
The segmentation based on the laser scanner data will be called 
the house-segment-file and the sublevel, the aspect based level, 
the roof-segment-file. Examples of both files are given in Figure 
3-3 and 3-5. The darker polygons are the map segments. 
Comparing Figure 3-2 with Figure 3-3, several houses, not on 
the map or ortho photo, are visualised and segmented. 
  
m am Em 
Figure 3:3. Example of house segments and map segments 
A- 171