International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
148 
The result of the whole feature extraction process is a symbolic 
description of the image content by segments (polygonal con 
tour lines) with a number of attributes (signature, structure, 
size, shape) and topological information (neighbourhood 
relations). Starting from the result of the segmentation (see Fig. 
9 as an example for the object class ’settlement’), a semantic 
modelling of the scene content is developed. 
35 
30 
•O 25 
c 
CO 
£•20 
w 
a 
■S 15 
>, 
« io 
Ü 
♦ Water 
^ Forest 
À Agriculture 
# Settlement 
‘»A 
A A 
10 15 20 
Grey values (band 4) 
25 
30 
Fig. 10. Modified standard deviation as texture parameter. 
250 
Ä 200 
TJ 
c 
CO 
.Q 150 
50 
♦ Settlement 
^ Forest 
A Agriculture 
# Water 
# 
A ^ A A, 
^ \ A A 1 
A ^ 
A A 
• 
♦ 
♦i* 
50 100 150 200 
Grey values (band 1) 
250 
Fig. 11. Texture parameter homogeneity. 
3. SEMANTIC MODELLING OF THE TOPOGRAPHIC 
OBJECTS 
The information from the segmentation and the topographic 
database ATKIS with their respective extracted features are 
stored in the object-relational database Postgres. Through this, 
the access on the data is simplified, especially because the used 
database is capable of storing and processing geometric data. 
By means of an intersection of both geometric scene 
descriptions from segmentation and ATKIS, we obtain a new 
unambiguous scene description with disjoint objects as 
explained in Section 3.1. The disjoint objects with their 
processed features are also stored in the relational database. 
This is our knowledge base for the next step of classification. 
The classification process is performed in a semantic network, 
which is able to process general and specific knowledge about 
the topographic and disjoint objects. This step is explained in 
section 3.2. After the classification, disjoint objects with the 
same semantic meaning and a common border are merged. The 
result is a complete semantic description of the scene. 
3.1. Building of Disjoint Objects 
Because the scene descriptions from DLM and segmentation 
have ambiguous semantics for certain areas (see Fig. 12), it is 
necessary to introduce a method to solve this ambiguity. To get 
an unambiguous scene description, an intersection between 
corresponding objects is performed. 
Fig. 12. Overlapping DLM and segmentation object. 
only segmentation 
only DLM DLM & Segmentation 
Fig. 13. Resulting disjoint objects. 
As an example (see Fig. 12 and Fig. 13), the intersection bet 
ween the corresponding DLM and segmentation objects is built. 
The result is a set of disjoint objects with three different 
semantics: the inner object where DLM and segmentation have 
the same semantics and the outer objects, where either the DLM 
or segmentation define the object as belonging to the object 
class ’settlement’. This is a simple example, because not only 
objects with the same object class may be overlayed. Of course, 
many DLM objects with a class other than the one of the 
intersecting image objects can be found. In addition, because of 
the classwise segmentation and digitizing errors, both 
geometric scene descriptions have overlapping objects from 
different classes. Therefore, the intersection process is not only