International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
in non-shadow areas. After the classification, the shadow and 
non-shadow pixels for each landuse class are combined again to 
obtain one unique class for each type of landuse. The final 
result of the classification algorithm for the whole test area is 
given in Figure 13. 
Fig. 12. Classification using shadow areas. 
Fig. 13. Classification result based on CIR imagery and laser 
scanning using predefined shadow areas. 
Our application of the classification is twofold. Firstly, it can be 
used for initial data capture; secondly, it is applied for 
verification of existing GIS. Since the manual revision of GIS 
data is very costly and time consuming, procedures are required, 
which enable a fully automated verification of existing GIS 
objects. Figure 14 shows an existing map of scale 1:5000 and 
the buildings classified by our approach, which are represented 
by the black polygons. At the upper left of the area, differences 
between the map and the result of the classification, which 
represents the actual situation, are clearly visible. Even though 
the result of the classification might not be detailed and precise 
enough especially for the automatic reconstruction of buildings, 
the automatic detection of inconsistencies can be applied to 
guide an operator-based revision of the GIS data. 
Fig. 14. Comparison of existing map using classification result. 
The second application we have in mind is the initial mapping 
of urban vegetation like tree regions. Single trees are usually not 
represented in standard GIS. However, they are required for 
tasks aiming at visualization or simulation in urban 
environments. For this type of applications, our classification 
result can be favorably exploited. 
4. DISCUSSION 
The paper demonstrated the benefit of combining laser data and 
colour aerial imagery for automatic scene labelling in an urban 
environment. The main advantage of our approach is that the 
problem of dealing with the mutual complementarity of the 
different data sources is solved implicitly by combining height 
and colour information in a classification step. The mutual 
weighting of the different data is represented by the cluster 
centres and the distances of the feature vectors to these cluster 
centres. Within a supervised classification, these parameters are 
provided for each channel and object class by the analysis of the 
training areas. 
Due to atmospheric effects, different spectral diffusion 
depending on the sunlight, different spectral characteristics of 
vegetation depending on season or soil, etc., new training areas 
have to be digitized for each dataset, if multispectral data is