x DATA 
glerstr. 7. 
tering process and 
ion of non-terrain 
fic modelling, like 
inning or disaster 
ilts. Therefore, the 
ight texture, shape 
1gh DTM (created 
algorithm on this 
st/last pulse height 
and a fuzzy logic 
ted in an overall 
(flat terrain). The 
s (4%). The most 
ing the statistical 
: will be extracted 
tion process. Two 
zzy logic and a 
"he influences and 
ons as well as a 
nt classification 
vestigations. 
erived exclusively 
tional information 
aused by specific 
nt - as mentioned 
ed out also during 
the other hand the 
nalysing airborne 
r in raster format 
re used, Karlsruhe 
. 2km x 2km) and 
nent, hilly terrain, 
ptured in first and 
additionally laser 
an subset of these 
dly permission of 
  
Figure 2. ‘Karlsruhe’ test area subset 
3. CLASSIFICATION OF 3D OBJECTS 
3.1 Definition of object classes 
As mentioned above, two test sites have been investigated, 
Salem and Karlsruhe. In this project, the most important aspect 
Was to investigate classification quality obtained by analysing 
laserscanning data with fuzzy logic methods. Therefore, the use 
of all main classes necessary for the applications defined above 
were included: buildings, vegetation and terrain. At test site 
Karlsruhe the amount of classes had to be restricted to buildings 
and vegetation because only one (man-made) terrain object 
occurs due to an extremely flat surface of the earth. 
3.2 Segmentation of 3D objects 
Although this approach analyses raster data not the commonly 
used pixel based classification was preferred but an object 
oriented method based on the segmentation of 3D objects.Some 
other works in this direction can be found e.g. in (Hofmann, 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
  
Figure 3. Segmented objects of ‘Salem’ test site 
  
Figure 4. Segmented objects of ‘Karlsruhe’ test site 
Maas, Streilein, 2002; Schiewe, 2001, Lohmann, 2002). In most 
cases the image processing system eCognition (Definiens, 
2001) is used. In opposite to these our approach is not based on 
general standard features but on the a-priori knowledge about 
the characteristic of the relevant 3D objects, i.e. about their 
specific appearance in laserscanning data (Voegtle, Steinle 
2003). 
In a first step of this approach a so-called normalised digital 
surface model (nDSM) is created to exclude the influence of 
topography (e.g. Schiewe 2001). For this purpose a rough 
filtering of the original laserscanning data (DSM) is performed 
to extract exclusively points on the ground (DTM). This 
filtering is based on our convex concave hull approach (von 
Hansen, Voegtle 1999) which results — by an accordant choose 
of the filter parameters - in a rough trend surface of the terrain 
(rough DTM) without vegetation or building points. Now the 
resulting nDSM is calculated by subtracting this DTM from the 
DSM. In this data set all 3D objects on the surface of the terrain 
remain, in some cases also a few terrain objects are included 
caused by rough rocks or sharp terrain edges. It is evident that 
this result hasn't to be perfect because non-relevant objects — in