Scanlines 
st shadows of buildings and trees reflect some of the infor- 
i+1 mation covered by a topographic map. For comparison the 
corresponding part of the map is plotted in figure 12. 
   
   
Original image 
      
  
  
   
Ss ann 
i+1 æ Flight path 
Projection 
centres 
Rectified image 
  
ge 
  
Z-plane 
X 
Grid point 
Ground point 
  
Figure 9: Image rectification 
4.3 Multispectral Analysis 
Our primary interest in extracting information from the mul- 
tispectral component of DPA is currently focused on the 
field of topographic and thematic mapping. Tasks like the 
extraction of roads or the detection and reconstruction of Figure 10: Near infrared channel 
buildings or other man-made objects are research topics 
of digital photogrammetry for which automatic procedures 
are under development. First successful experiments have 
been presented by Haala and Hahn (1995) and Weidner 
and Fórstner (1995). Major information source used in that 
work is the digital terrain model and the aerial stereo image 
pair. Critical steps of those procedures are the detection 
  
  
  
  
  
of objects and the discrimination of different object classes. 
With respect to this critical steps we expect that the mul- + 
tispectral data of DPA will be a very useful source. This 
tified assessment is supported by the work of Shettigara et al. 
(1995). They developed a procedure for extraction of man- i 
made objects from multispectral aerial images. s. 
For the multispectral investigation an area is selected in 
which vegetation, buildings, streets and other objects are 
f the imaged. The near infrared channel of this area is depicted 
jane in figure 10. 
ane). For the classification we used the isodata algorithm. This 
d on is a well established iterative optimization clustering proce- 
n fig- dure. It is based upon estimating some reasonable asign- 
A the ment of the pixel vectors into candidate clusters and then 
ulting moving them from one cluster to another in such a way that 
y the the sum of squared error over all pixels to the correspond- 
nsfer ing cluster mean is reduced. To get a reasonable assign- | vegetation 
jhted ment for candidate clusters supervised classification with Roof 
within an upper limit for the number of classes is used. For exam- 00 
one ple, we have chosen a maximum of 30 classes. Streets 
case The result of the Maximum Likelihood classification to- Shadows 
is as- gether with the labeling of the classes is shown in figure 
hown 11. This first result with the very coarse classes of vege- 
tation, roofs (buildings), roads or other sealed areas and Figure 11: Multispectral classification 
145 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B2. Vienna 1996