The generic model was obtained by discussing the manual delineation of two experienced 
cartographers. 
In Fig. LI a problem solving KB is displayed capturing the generic model. Control is 
passed to this knowledge base by sending an instantiate-road-network message to an actual 
member of the landcover class (cfr. REG1 in Fig. Bl). 
• First, an initial network is obtained by tracking a thinned thresholded result of 
applying a line detector (currently, we use the TON operator, [6]) to the satellite 
image within a landcover region of interest. This network is represented by members 
of the LINE and the CROSSING classes. A LINE instance represents a linear element 
having high photometric evidence to be a part of a road. A CROSSING corresponds 
to the crossing of at least two lines. 
• Because the generic model stresses the importance of rectilinear groups of lines, a 
RELATION subclass STRUCTURAL is included. It has itself subclasses to describe 
L-shaped, U-shaped and rectangle structures. Therefore, in a second phase, simple 
heuristics are used to obtain related objects : for example, a line connecting two 
crossings will generate a U-structure if at both crossings lines can be found, parallel 
to one another and on the same side of the former : as a result a new member of 
the class U-STRUCTURE will be created having a related-objects slot containing 
the three lines. A U-structure is completed to a rectangle, if the missing crossings 
and lines can be found or justified from the primal network. Optimal outlines of 
the rectangle sides are obtained by F*-optimization on the result of the line-operator 
image. 
• When a rectangle is found, it is used in combination with a U-structure or a L- 
structure to complete the latter to a rectangle : from the rectangle, a rectangular 
search environment is constructed and overlayed at the corresponding corner(s) of the 
L or the U (see Fig. L2); within this context-derived template, evidence for missing 
lines and crossings is obtained by examining predicted crossing-positions and the 
existing road network in this area. Sufficient evidence is determined heuristically 
and depends on parameters such as length and slope of lines, percentage overlap of 
lines with the template region, possible crossing of lines, distance between a crossing 
and a predicted position, ... . If acceptable evidence is found, the optimal outlines 
of the new rectangles again result from /^-optimization 
• Roads that are barely detectable on the image photometry can be suggested by 
partial matches of the templates to given U’s and L’s. It is left to the user to accept 
or reject these structurally predicted roads. 
Fig. L4 shows a result obtained on the SPOT image from Fig. L3. Fig. L6 shows the 
result on the Landsat TM image from Fig. L5. 
652