International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
layer. Topological relations provide information about the kind 
and the properties of neighboured objects. Therefore, the class of 
attributed relations (attr-rel) is introduced. In contrast to other 
relations, this one has attributes which can be used to constrain 
the properties of the connected nodes. For example, a topological 
relation close-to can be generated which restricts the position of 
an object to its immediate neighbourhood. The initial concepts 
which can be extracted directly from the data are connected via 
the data-of link to the primitives segmented by image processing 
algorithms. 
For the efficient representation of multiple relations, the 
minimum and maximum number of edges can be defined in the 
knowledge base. The minimum quantity describes the number of 
obligatory relations and the difference to the maximum quantity 
represents the number of optional relations between objects. In 
this way, it can be easily modelled that for example a crossroad 
consists of three up to five intersecting roads. Additionally, for 
each edge a priority can be defined in order to realize an ordered 
evaluation of the relations. Edges with high priority are 
instantiated first. For the application of landscape analysis for 
example, it can be guaranteed that the streets are extracted prior 
to the rivers. 
Some relations appear exclusively in certain domains. For 
example roads have always a lane but they have pavements in 
urban areas only. This fact is taken into consideration by a 
domain dependent relation in the generic model. Fig. 1 shows a 
simple semantic net for a generic model of a Road Net which is 
defined as a composition of at least one Road, illustrated by the 
set [1,«]. A Road consists of one or two lanes. Its specialization 
Major Road inherits the properties of Road and possesses an 
additional Crash Barrier. For the part-of relation between 
pavement and road the domain Urban Scene is defined. Only in 
urban scenes this relation is valid and the system searches for 
pavements. All the initial objects Crash Barrier, Lane, and 
Pavement are represented by a Stripe-Form in the image. 
Figure 1. Example for a semantic net: The scene contains at least one 
Road. The Pavement is defined for the domain Urban 
Scene. The more special concept Major Road inherits the 
properties of Road. All objects are represented by a 
Stripe-Form in the image. 
2.2. Control of the Scene Analysis 
To make use of the knowledge represented in the semantic net 
control knowledge is required that states how and in which order 
scene analysis has to proceed. The control knowledge is 
represented explicitly by a set of rules. The rule for instantiation 
for example changes the state of an instance from hypothesis to 
complete instance, if all subnodes, which are defined as 
obligatory in the concept net, have been instantiated completely. 
If an obligatory subnode could not be detected, the parent node 
becomes a missing instance. 
An inference engine determines the sequence of rule execution 
according to a given strategy. A strategy contains a set of rules out 
of the rule base. For each valid rule, a priority is defined to 
determine in which order the rules are tested. The first matching 
rule is fired. The user can modify the interpretation strategy by 
changing the priorities and by removing or inserting rules to the 
current strategy. The default strategy prefers a model-driven 
interpretation with a data-driven verification of hypotheses. 
Topological relations are instantiated as soon as possible to 
realize a spatial reasoning. 
Whenever ambiguous interpretations occur, for example if more 
than one suitable image primitive is found for a hypothesis, they 
are treated as competing alternatives and stored in the leaf nodes 
of a search tree. Each alternative is judged by comparing the 
measured object properties with the expected ones. The 
judgement calculus models imprecision by fuzzy sets and 
considers uncertainties by distinguishing the degrees of 
necessity and possibility (Dubois, 1988; Tonjes, 1999). The 
judgements of attributes and nodes are fused to a judgement of 
the whole interpretation. The best judged alternative is selected 
for further investigation. 
Starting at the root node of the concept net, the system generates 
model-driven hypotheses for scene objects and verifies them 
consecutively in the data. Expectations about scene objects are 
translated into expected properties of the corresponding image 
primitives to be extracted from the sensor data. Suitable image 
processing algorithms are activated and the semantic net assigns 
a semantic meaning to the returned primitives in a data-driven 
way. Interpretation stops, if a given goal concept is instantiated 
completely or no further rule of the current strategy can be fired. 
3. KNOWLEDGE BASE FOR THE INTERPRETATION 
OF REMOTE SENSING IMAGERY 
For object extraction, only those features are relevant that can be 
observed by the sensor and that give a hint for the presence of the 
object to be extracted. Hence, the knowledge base contains only 
the necessary and visible object classes and properties. The 
network language described in chapter 2.1. is used to represent 
the prior knowledge by a semantic net. In Figure 2 a generic 
model for the interpretation of remote sensing images is shown. It 
is divided into the 3D scene domain and the 2D image domain. 
The 3D scene domain splits into the semantic layer and the 
physical layer. If a geoinformation system (GIS) is available and 
applicable, an additional GIS layer can be defined representing