International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
landuse 
poor 
landuse 
with 
are two 
of 
of an 
scene 
and 
is new. 
of this 
spectral 
time 
Therefore, a model-driven top-down approach can be 
integrated into the common data-driven bottom-up process of 
satellite image analysis. Figure 1 shows the flowchart of the 
analysis process. Common satellite image analysis 
techniques are restricted to pixel-based classification and 
involve the use of only one feature (spectral signature) and 
interactive selection of training areas. With this new 
approach these restrictions are avoided. 
Image and topographic databases give a description of the 
scene from their respective perspectives. By using the a 
priori semantic information of the topographic objects in the 
map, an automated selection of training areas is performed. 
This is done by overlaying the image with the topographic 
objects. The geometric errors between the image and map 
objects are eliminated through the large amount of training 
areas and the use of a histogram analysis. The learning of 
typical features for the object classes is necessary for the later 
step of classification. 
In the following step involving the semantic modelling of the 
topographic objects, both symbolic scene descriptions are 
linked and an unambiguous scene description with disjoint 
objects is built. Knowledge based techniques are applied for 
the classification of the resulting geometric disjoint objects. 
The result of the classification process is a complete semantic 
scene description. 
This paper does not deal with the last step of updating the 
digital database, which involves its comparison with the 
semantic scene description. 
2. KNOWLEDGE BASED FEATURE EXTRACTION 
AND SEGMENTATION 
The common features in satellite image analysis, i.e. the 
spectral signatures (mean values, standard deviations), have 
been proven to be insufficient for high quality results (Bähr 
and Vögtle, 1991; Vögtle and Schilling, 1995). Therefore, 
these features have to be extended to spectral as well as non 
spectral parameters, which can contribute to an improved 
distinction between the defined object classes. Thus, 
geometrical and structural features are taken into account 
(Table 1): 
Spectral features 
Spectral Signature 
Texture 
Non-spectral features 
Structure 
Size 
Shape/Contour 
Neighbourhood Relations 
Table 1. Selected features for image analysis. 
The automated extraction of the above defined features is 
based on the a priori knowledge represented in the 
topographic database ATKIS-DLM200, which offers both a 
(possibly not up-to-date) geometric and semantic description 
of those landuse objects to be extracted from satellite images. 
In contrast to the commonly used method, where a human 
operator interactively has to define some representative 
training areas based on his experience and intuition, now all 
DLM-objects of the same class within the geocoded satellite 
image can be used as training areas without human 
interaction. Therefore, a very large sample is taken and a 
robust estimation of the defined features is performed to 
exclude disturbances caused by errors in the topographic 
database or in the image information, e.g. out-of-date status 
of some polygons (contour lines), digitizing errors or errors 
in the geometric correction of the satellite image. For a 
robust estimation, it is assumed that for each class in the 
image at least more than 50% of the underlying DLM object 
area belongs to the DLM class, a condition which is fulfilled 
in most cases. 
The feature extraction process in this project contains a 
hierarchical concept. The spectral characteristics of the 
objects is still one of the most important features in satellite 
image information. To get a robust estimation of the spectral 
signatures of each object class, only the representative 
reflectance values for this class are extracted. For relatively 
homogeneous objects, like 'water' or 'forest', statistical 
methods have been proven to be sufficient, e.g. histogram 
analysis (extraction of the standard deviation) or a median 
estimation. Inhomogeneous objects, like 'settlement areas', 
can not be treated in this way. Typically, these areas contain 
a strong mixture of different (sub-)objects (man-made 
objects, meadows, gardens, trees, water areas etc.), and 
therefore, a wide range of reflectance values. Nevertheless, 
the accumulation of vegetation-free pixels caused by man 
made objects (e.g. buildings and traffic areas) can be seen as 
representative for 'settlement'. With respect to this 
knowledge, vegetation-free pixels can be extracted, by means 
of the NDVI (Normalized Difference Vegetation Index): 
NDVI = 
(IR-R) 
(ir + r) 
IR ...reflectance values in near infrared 
R ...reflectance values in visible red 
In Fig. 2, the NDVI for different topographic classes is 
shown. 
Fig. 2. Normalized Difference Vegetation Index (NDVI).