anbul 2004 
  
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004 
  
The calculation of the percentage of sealed ground is done in 
the same way as the calculation of the percentage of trees with 
the difference that all street pixels are selected from the pixel- 
based classification result (Figure 7 d). Industrial settlement 
objects have typically a higher percentage of sealed ground and 
are represented by bright grey values whereas residential 
settlement objects have typically a lower percentage of sealed 
ground and are represented with darker grey values. 
The calculation of the textural appearance is shown in 
Figure 7 e. We use the same texture operator as described in 
section 3.2.2 and calculate the average texture per object. 
Industrial settlement objects are represented typically with 
brighter grey values because they have a more homogenous 
textural appearance whereas residential areas are represented 
typically with darker grey values because they have a more 
inhomogeneous textural appearance. 
3.4.3 Classification: The five object characteristics span a 5- 
dimensional feature space. The feature vector of each object is 
evaluated and classified either to the object class residential 
settlement objects or industrial settlement objects. Table 1 
shows the feature vectors of some objects of the test site. All 
values are mapped onto the interval [0..255], like in a typical 
pixel-based classification. 
Table 1. Feature vectors of the objects 
  
  
  
  
  
  
object average | average | percent. percent. average 
number house | roof slope trees sealed texture 
size ground 
AO01BHSD 03 85 20 99 55 
A01BHS85 25 29 8 34 102 
A01BH86 25 56 11 76 61 
AO01BH7W 185 65 21 86 67 
A0IBH7X 191 48 7 34 50 
  
  
  
  
  
  
  
  
  
In order to classify the objects, we use a maximum likelihood 
approach that classifies the objects into the most likely class 
based on the evaluation of training data. The objects of the 
existing GIS database represent the training data. 
Figure 8 a shows the distribution of the percentage of trees for 
all objects of the GIS database and Figure 8 b the distribution of 
the percentage of tress for industrial settlement objects and 
residential settlement objects. It can be seen in the diagram that 
the higher the percentage of trees the higher is the likelihood 
that an object is representing a residential settlement area. But it 
can also be seen that there is an overlapping area where an 
object can represent a residential settlement object or an 
industrial settlement object with a similar likelihood. 
Ín order to make the distinction between the two object classes 
clearer, we evaluate not only one object characteristic but more 
than one. Figure 9 a shows in a scatterplot the two-dimensional 
distribution of the two object characteristics ‘percentage trees’ 
and ‘average house size’ for all objects in the GIS database. 
Figure 9 b shows the same distribution only for residential 
settlement objects and Figure 9c for industrial settlement 
objects. It can be seen that the distinction between the two 
object classes becomes clearer. The more object characteristics 
are evaluated the clearer becomes the distinction. All object 
characteristics are used in the object-based classification. That 
means that we evaluate a 5-dimensional feature space and 
classify the objects to the most likely class based on the 
statistical distribution of the training data. 
  
   
a) 
frequency 
percentage tree 
industrial settlement objects 
bj 
  
percentage trees 
Figure 8. Distribution of percentage of trees 
  
  
  
  
  
  
  
M s : 3 
PA: : E pas à 
« J Fr * Es > 5 P % Tis > * 
3 s" V WM uate : ex 4 Th ms x En + - 
a) b) c) 
Figure 9. Scatterplot (x-axis: percentage trees, y-axis: 
average house size) 
4. RESULTS AND DISCUSSION 
The approach was tested on a test area with 24 km” that contains 
190 residential settlement objects and 84 industrial settlement 
objects. The test site is Vaihingen/Enz that is located in the 
southern part of Germany and represents a rural environment 
with smaller settlements. The multispectral data were captured 
with the DMC camera system, which is a CCD-matrix based 
camera system with 4 multispectral channels: R, G, B and Near 
Infrared (Hinz 2001). The LIDAR data were captured with the 
TopScan system and have an average point distance of 
approximately | m (Schleyer 2001). The LIDAR data and the 
multispectral data were resampled into regular raster images 
with a pixel size of Im. The tests were carried out with ATKIS 
datasets. ATKIS is the German national topographic and 
cartographic database and captures the landscape in the scale 
1:25,000 (ADV 1988). 
In a manual classification all residential and industrial 
settlement objects of the databases were compared with the 
images and subdivided into the classes OK, unclear and not OK 
(see Figure 10). The class OK contains all objects with no 
change in the landscape (172 + 64 = 236 objects). The class 
unclear contains all objects where it was unclear if there was a 
change or not without evaluating additional sources (18 + 19 = 
37 objects). The reason for the relatively high number of objects 
in that class is that the distinction between residential and 
industrial objects in ATKIS is not only done because of the 
spectral appearance but also because of non-visible criteria’s 
(for example *non disturbing trade and repair businesses" have