650 
of forest paths has been obtained by interaction with human experts. The model states that 
such paths are straight, show perpendicular crossings and appear as repeating geometrical 
structures (rectangles). Usually chains of roads or paths near the border of a forest region 
can be expected. This model is used to aggregate the results of line extraction into a 
network description. Knowledge employed is based on rules of perceptual grouping and of 
structural recognition. A similar model is used to extract roads from desert areas that are 
being transformed into cultivated land. 
The second case study exploits the influence of the terrain on the appearance of roads in 
mountainous area. Important terrain-related characteristics of roads such as maximal slope 
and the tendency to follow ground contours are used to validate line elements obtained 
by photometric analysis. Ground contours are used as guide-lines that can be elastically 
deformed under influence of photometric force-fields to obtain optimal road outlines. 
The third case study investigates the usefulness of hydrography to road extraction. A 
binary hydrographic map is transformed into a hierarchical description of the river network, 
according to principles of physical geography. This description is used as an index in the 
image data to find and validate line structures as roads. Rivers are transformed into search 
environments to perform local line extraction. A search environment depends on the order 
of the river and on the terrain model near that river. 
Existing road information is used as a logical framework in all the case studies, because 
existing roads are often invisible due to tree canopy closure. 
2 Background 
In [7], we have overviewed previous work on road extraction from satellite imagery. Two 
main knowledge sources are mentioned in these papers : first, the image photometry and 
basic models to find lines, second, rules of perceptual grouping used to select, to combine 
and to extend line primitives according to models of how roads appear in a landscape. Two 
remarks can be made : 
• In general, this knowledge is shallow as it is not (or weakly) based on any theoret 
ical model from (physical) geography or landscape investigation that explains the 
possibility to find roads in particular areas. 
• The domain of applicability of previous attempts is not specified or seems to be very 
restricted. 
We advance the hypothesis that additional data and knowledge are needed to overcome 
these shortcomings. We also detailed in [7] the knowledge sources mentioned in 1 that are 
used by human analysts during road delineation. Briefly, this discussion can be summarized 
as follows : 
• Different types of landcover correspond to different types of road network topology. 
These generic models imply typical ways to instantiate road network extraction.