International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
Since the use of geometric information during multispectral 
classification is the main purpose of the algorithm, the 
acquisition, processing and analysis of the height data is of 
major interest. For this reason, these steps will be described in 
detail in the following section. 
2. GEOMETRIC INFORMATION 
2.1. Acquisition of Height Data 
The geometric information, which is used during the 
classification step, has to be extracted from a DSM. This DSM 
can be obtained by automatic image matching algorithms 
applying stereo aerial imagery or can be directly captured by 
airborne scanning laser systems. These airborne laser scanning 
systems are capable of making dense and well-distributed three- 
dimensional measurements on the terrain surface. As in 
tachymetric data acquisition, the coordinates of terrain points 
are determined by polar measurement. For point determination, 
the run-time of a laser pulse reflected at the ground is used to 
calculate the distance between the sensor and the laser footprint, 
whereas position and orientation of the sensor is determined by 
an integrated GPS/INS system. With current systems, terrain 
points can be measured at approximately one point each 0.5 x 
0.5 m 2 with an accuracy in the order of 0.3 m (Lohr (1997)). An 
example of a DSM acquired by laser scanning is shown in 
Figure 1. In order to improve the 3D visualisation of the data, 
the ortho-image has been overlayed on the DSM surface. 
Fig. 1. 3D view of test area, height data obtained from laser 
scanning. 
In principle, image matching can alternatively be applied for 
height data acquisition. The main advantage of this approach is 
that the same sensor can be used to provide the required 
geometric and radiometric information. Image matching 
techniques have become standard tools for three-dimensional 
surface acquisition in open terrain. However, they suffer from 
problems in built-up areas due to occlusions and height 
discontinuities. In these areas, the DSM quality mainly depends 
on the presence of texture at roof regions and on the amount of 
contrast between roof and terrain surface (Price and Huertas 
(1992)) This results in considerably different DSM quality at 
roof regions, even in the same image pair. In contrast to that, 
direct measurement by airborne laser scanners usually provides 
DSM data of higher and more homogeneous quality, especially 
in urban areas. This was the reason for using this technique in 
this application. 
2.2. Preparation of Height Data 
In the example dataset depicted in Figure 1, it is clearly visible 
that DSM not only represent the terrain surface like Digital 
Terrain Models (DTM), but also contain buildings and other 
objects like trees, which are higher than their surroundings. To 
make the information on these objects accessible, the so-called 
normalised DSM, i.e. the difference between DSM and DTM 
has to be calculated as the first step. This results in a 
representation of all objects rising from the terrain. The required 
DTM can be derived from the measured DSM by mathematical 
grey level morphology as suggested by Weidner and Forstner 
(1995). In their approach, the DSM is processed by 
morphological erosion, which is followed by morphological 
dilation. The combination of erosion and dilation results in an 
opening of the DSM surface, eliminating all local maxima in 
height of a predefined size. This predefined size is for example 
based on the expected maximum extent of a building and is 
used a priori in order to define the size of the morphological 
operator to be applied. After morphological processing, the 
resulting surface roughly represents the terrain surface, i.e. an 
approximate DTM is obtained. 
Depending on the aspired application and the available data 
sources, existing GIS data can be also used in combination with 
the measured DSM for the approximate reconstruction of the 
terrain surface. For DTM generation, height values at terrain 
regions have to be derived from the DSM. If for example two- 
dimensional representations of streets are available from a 
standard database, height values at these regions can be 
extracted from the DSM and utilized to generate the required 
surface by interpolating between these areas. In our application, 
ATKIS has been exemplary used for that purpose. ATKIS is the 
German topographic cartographic database and presently 
contains more than 60 different feature types for the whole area 
of Germany in the scale 1:25,000. Within this database, roads 
are represented by linear objects, thus the required information 
is available at sufficient accuracy. 
2.3. Object Detection 
DSM have already been used for the automatic detection of 
buildings to trigger their subsequent geometric reconstruction 
from stereo image data (Baltsavias, Mason and Stallmann 
(1995)). Even though objects rising from the terrain can be 
detected quite well from the height data, the discrimination 
between buildings and trees can be difficult, if only simple 
criteria like region size or shape are considered. A possible 
approach is to use the roughness of the DSM surface measured 
by differential geometric quantities as an additional criterion for 
the discrimination of buildings and vegetation (Brunn and 
Weidner (1997)). However, due to the restriction to surface 
geometry, the number of object types, which can be 
discriminated within a DSM is very limited.