653 
3.2 Exploiting DTM knowledge 
In the second case study, we use a Digital Terrain Model (DTM) to guide road extraction. 
In particular, we are interested in the influence of the terrain on the appearance of roads 
in mountainous areas. Generally, these roads follow contours in order to minimize con 
struction and travelling effort. The following is a generic model obtained for delineating 
mountain-roads : 
• in general, mountain-roads are serpentine 
• mountain-roads usually follow ground contours 
• if contours are not followed, then altitude differences are minimized 
• the slope along the track of a road has a maximal value 
A part of a mountainous country SPOT image and the result of applying the TON 
detector to it are shown in Fig. D1 and Fig. D2, respectively. From the latter, line seeds 
having high response are chosen. If a seed has a number of points of high line curvature, 
it is assummed to be part of a sinuous road. From the DTM an average contour line 
connecting the individual pixels of the seed is obtained. This is illustrated by the white 
line on Fig. D3. The DTM image itself was obtained by TIN-interpolating contour lines, 
manually digitized from an existing topographic map. The contour line obtained in this 
way is used as a deformable model to extend the seed element according to the second 
statement from the generic model. This iconic deformation procedure inherits from a 
general method known as snakes, see [2]. 
By elastically deforming the contour line under influence of an external photometric 
force-field and of an internal energy, an optimal road outline is obtained. The force-field 
used in this case is the magnitude response from the line detector and displacements in 
this field are guided by the detector’s directional response. Our development tool offers 
flexible ways to tune parameters and gather data about the line’s movement. A result of 
this method is shown in Fig. D4, where part of the contour line from Fig. D3 has been 
deformed to the road outline. Work is under way to include the DTM itself (and not just a 
contour line out of it) in the optimization process. Also the combination of the serpentine 
lines into a network is under investigation. 
3.3 Exploiting HYDROGRAPHIC knowledge 
In the third case study, a hydrographic map is used as an index in the image data to look 
for roads almost parallel to rivers. The position of a road in a terrain near a river is related 
to the order of that river in the hierarchical drainage network it is part of. Again, a generic 
model was obtained from discussion with cartographers : 
• roads are usually parallel to contours tracing the valley of the river rather than to 
the river itself