International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
Another possibility to improve object detection based on DSM 
resulting from airborne laser scanning is the further analysis of 
the reflected laser beam. Laser scanning systems can be 
separated into continuous wave or pulsed laser systems. If a 
pulsed laser system is applied, multiple reflections will occur 
during the acquisition of trees. As depicted in Figure 2, during 
measurement of trees a certain percentage of the laser footprint 
will be reflected by the branches and leaves of the tree. Other 
parts will penetrate the foliage and will be finally reflected by 
the terrain surface. For this reason the top of the tree refers to 
the first echo of the laser pulse, which is recorded by the laser 
sensor, while the last echo usually refers to the terrain surface. 
first response 
reflections at 
foliage 
last response 
reflection at 
terrain 
Fig. 2. Reflection of a laser pulse at trees. 
Fig. 3. Grey value representation of DSM derived from first 
echo measurement. 
If the laser system is capable of recording and discriminating 
multiple laser pulse echoes, they can be utilized in order to 
separate trees and buildings. Figure 3 shows a grey value 
representation of a normalized laser DSM. The original DSM, 
which was already depicted as 3D visualization in Figure 1, is 
based on the first echo measurement. For this reason, both trees 
and buildings are visible. Figure 4 shows the corresponding 
result for a DSM derived from last echo measurements. In this 
example, only the buildings are visible. Hence, the difference 
between first and last echo normalized DSMs can be used for 
the detection of tree regions. 
Fig. 4. Grey value representation of DSM derived from last 
echo measurement. 
The laser system we are using for DSM acquisition is not 
capable of simultaneous recording of multiple echoes. 
Currently, either the first or the last reflection of the emitted 
laser pulse can be measured. Since this prevents the 
measurement of the required data in a single pass coverage, the 
flight effort for laser data capture is doubled, if one aims at the 
acquisition of the first and last response of the emitted laser 
pulse. Additionally, in our examples for some areas no response 
could be measured at all in the last pulse mode. These regions 
correspond to the white areas depicted in Figure 4. Besides 
these sensor-related problems, a further differentiation of object 
classes like the extraction of streets or different landuse classes 
like grass-covered areas is not possible, if only laser data is 
applied. For this reason, in our approach the height data is 
integrated with multispectral imagery within a combined 
classification step in order to separate the required objects. 
3. CLASSIFICATION OF URBAN AREAS 
3.1. Spectral Data 
For the test site, color infrared (CIR) aerial images were 
available, which were taken at a scale of 1:5000 with a normal 
angle aerial camera. For digitization, the images were scanned 
at a resolution of 60 p.m, resulting in three digital images in the 
spectral bands near infrared, red and green with a pixel footprint 
of 30 cm. The basic idea of the proposed algorithm is to 
simultaneously use geometric and radiometric information by 
applying a pixel-based classification. Within this classification, 
the normalized DSM is used as an additional channel in 
combination with the three spectral bands. For integration of 
different data types, the first problem to be solved is the 
registration of the datasets. In order to transform the data a 
common system a colored ortho-image is generated from the