International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
6 7 8| 9 10 
Fig. 3 a. False color composite of NIR (red), red (green) and green (blue) bands; b. ISODATA clustering with 6 classes; c. ISODATA 
clustering with 10 classes (weak boundaries between classes 4 and 5, 6 and 7, and 8 and 9). For explanation of the classes 
see text. 
Fig. 4. Edges extracted from the visible (blue), NIR (green) 
and thermal (red) multispectral images were combined 
into a color composite and then superimposed on the 
aerial photograph. The colors indicate which spectral 
image has the strongest discontinuity along the edge. 
4.2. Laser scanning data and aerial imagery 
Laser scanning data. Laser scanning systems are 
increasingly being used in photogrammetry, mainly for 
generating DEMs. Applications are as diverse as determining 
the topographic surface and the canopy of forested areas or 
establishing city models. Usually, the laser system is the only 
sensor used on the platform. This limits the range of problems 
that can be solved. More complex applications require several 
sensors to be used in concert. As briefly described in section 
3, the test site in Ocean City includes laser scanner data, 
aerial, and multipectral imagery. 
Laser scanning systems provide a fairly dense set of points on 
the surface. The accuracy in elevation is about 1 dm and 
footprint sizes are 1 m or less. The platform orientation 
system determines the positional accuracy. The critical 
component is the attitude. While the errors resulting from 
GPS and ranging are virtually independent of the flying 
height, the attitude error propagates linearly and thus restricts 
the flying height. Current airborne laser systems hardly 
exceed flying heights of 2000 m. 
Refined and segmented surface. In our attempt to recognize 
objects from multisensor data, the information the laser 
system provides is used for surface reconstruction and 
generation of hypotheses of man-made objects, such as 
buildings. It has long been realized that surfaces play an 
important role in object recognition. The laser points are not 
directly suitable for representing the surface. For the purpose 
of object recognition, we need an explicit description of 
surface properties such as breaklines and surface patches that 
can be analytically described. We distinguish between the 
raw, refined, and the segmented surface (Schenk, 1995). The 
irregularly distributed laser points describe the raw surface. 
The refined surface includes surface information obtained 
from aerial imagery. It is the result of fusing features from the 
two sensors. The fusion process also includes the resolution 
of conflicts that may exist between the laser surface and the 
visible surface. The next step is concerned with segmenting 
the refined surface, resulting in an explicit description that is 
much more amenable for object recognition than the raw 
surface, where important surface properties are only implicit.