651 
• A terrain model is useful to indicate plausible road tracks in the image data and 
provides constraints related to admissible slopes of lines. 
• The hydrography has obvious influences on the appearance of road structures. 
Roads often follow countour lines in valleys while minimizing the number of river 
crossings. 
• An existing roadmap can be used as a guide for image analysis to find analoguous 
roads and as a logical framework : new roads are usually connected to existing ones. 
We regard a GIS as a digital raster database consisting of layers of cartographic maps 
and RS data. It is assumed that the different data layers are registered in such a way 
that any two of them can be exactly (pixel-by-pixel) superimposed. To provide a smooth 
link between GIS and knowledge-based programming, an object-oriented environment for 
image understanding is used, [1]. It is implemented on top of the existing hybrid tool KEE, 
see e.g. [5]. Data sets from the GIS can be directly mapped to object classes in a KEE 
knowledge base (see Fig. Bl) (in KEE a set of related objects is called a knowledge base 
or KB). Possibilities to utilize iconic structures are provided by the image understanding 
environment. Expertise about road delineation is added as shown in the next sections. 
3 GIS-guided extraction : three case studies 
Because of the inherent ambiguity of satellite images and because a variety of road networks 
does exist, each showing different scales of complexity, there exists a need for a problem 
solving strategy that progressively reduces the search space. To cope with this problem 
we start from generic models of road-network appearance and try to incrementally focus 
the attention in order to obtain optimal results (with regards to some objective function). 
This strategy has been called hierarchical search, [4]. 
3.1 Exploiting LANDCOVER knowledge 
In the first case study, we deal with a generic network model of road appearances that has 
the following characteristics : 
• the network consists of straight roads 
• in general, these roads intersect perpendicularly 
• the landcover region is subdivided into repeating geometrical structures (such as 
rectangles) by the roads 
Typical landcover regions that show this kind of network patterns are forests in flat 
terrain (see Fig. L3) or desert areas being cultivated by irrigation projects (see Fig. L5). 
On high resolution satellite imagery, these roads are visible as line-like structures. The road 
network topology in such areas is related to practical considerations (e.g., accessibility).