International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
150 
this concept does not depend on other concepts, it corresponds 
to a disjoint object and its attributes can be fetched. The 
assignment of concrete attributes to a concept is called 
instantiation. 
The analysis now moves bottom-up to the concept at the next 
higher hierarchical level. If instances have been found for all 
parts of this concept, the concept itself can be instantiated. 
Otherwise, the analysis continues with the next concept not yet 
instantiated on a lower level. After an instantiation, the 
acquired knowledge is propagated bottom-up and top-down to 
impose constraints and restrict the search space. Thus, in the 
analysis process, top-down and bottom-up processing 
alternates. 
Generally, while performing an instantiation, it is possible to 
establish several correspondences between a concept and an 
object. However, only one of these correspondences leads to 
the correct interpretation. Usually, it is not possible to reliably 
decide at the lower levels which correspondence is correct; 
hence, all possible correspondences have to be taken into 
account. 
Thus, the image analysis is a search process, which can be 
represented graphically by a tree structure. Each node of the 
tree represents a state of the analysis process. If several 
correspondences are possible in a given state, the search tree is 
splitted. The analysis process continues with that node of the 
search tree, which is considered to be the best according to a 
problem dependent valuation. The problem of finding an 
optimal path in a search tree can be solved by the A*-algorithm. 
For further explanation of this tree search algorithm see Nilsson 
(1971), Dreyfus and Law (1977) or Kummert (1992). Its 
application is possible, if the path from the root node to the 
current node can be evaluated and if an estimation can be given 
for the valuation of the path from the current node to the 
terminal node. 
The functions, which evaluate the states of the analysis, are 
very important, since they are not only responsible for the 
efficiency of the search, but are also decisive for the success or 
failure of the analysis. The valuation of the search path is 
related to the valuation of the analysis goal. The valuation of 
the goal is calculated by considering the valuations of the 
instances and modified concepts, which are created in the path 
from the current node to the solution node. When an 
instantiation is performed, a hypothesis of match is implicitly 
established between the concept under instantiation and the 
chosen primitives (disjoint objects) from the database. 
The computed valuations for the instances and modified 
concepts in each state of the analysis are measures of our 
subjective belief in these hypotheses. They take values between 
0 and 1 and can be interpreted as basic belief values in the 
framework of the Dempster-Shafer theory (Dempster, 1967; 
Dempster, 1968; Shafer, 1976; Smets, 1991). The higher a 
valuation, the stronger our subjective belief in the 
corresponding hypothesis. The different valuations are com 
bined and propagated in the hierarchy of the semantic network 
resulting thus in the valuation of the analysis goal. 
Two aspects are evaluated for our hypotheses of a match: the 
compatibility and the fidelity of the model. The compatibility 
evaluates an analysis state considering the principles of 
perceptual grouping. It is calculated based on the geometrical, 
topological and radiometric properties of the disjoint object 
primitives. The compatibility can be seen as a measure of the 
ability to form an object of the generic model with the chosen 
image primitives. The fidelity of the model determines the 
quality of fit between the attributes of image primitives and the 
statistically learned attributes from the analysis of the DLM- 
information. Model fidelity is a measure of the ability to form 
exactly that object, which is predicated by ATKIS. 
3.3. Result of the Semantic Classification 
The result of the semantic network analysis is a new knowledge 
base of classified disjoint objects. Each object belongs to one of 
the four basic object classes but describe only a small part of 
the whole topographic object. Therefore, disjoint objects with 
the same semantic meaning and a common border are merged. 
The result is a complete semantic description of the scene. 
4. CONCLUSIONS 
There is a basic need for techniques to analyse image data in an 
automated way. With the here presented method, it is possible 
to get good results without human interaction. Many 
refinements have to be implemented until this goal is achieved. 
Experiences with an extended feature base and a special 
segmentation process confirm the efficiency of our concept by 
leading to a better separability of object classes. More suitable 
features should be found in order to increase the knowledge 
base used in the semantic classification process. Until now, a 
comparison between the presented segmentation process and 
classical processes is missing. The determination of good 
valuation functions for spectral as well as non-spectral features 
in the decision process that is performed in the semantic 
network has been proven to be a very complex and time 
consuming task. The structure of the semantic net is still 
simple, and can be extended in an easy way by adding new 
concepts and links. A semantic network for the classification is 
one knowledge representation among many others and its 
potential has to be verified with additional investigations. 
ACKNOWLEDGEMENTS 
This project is funded by the DFG (German Research 
Foundation) II C 5 - BA 686/10-3. 
REFERENCES 
AdV: Amtliches Topographisch-Kartographisches 
Informationssystem (ATKIS), 1989. Arbeitsgemeinschaft der 
Vermessungsverwaltungen der Länder der Bundesrepublik 
Deutschland (AdV), Hannover, Germany.