654 
• less important rivers (having a lower order) show narrow V-shaped valleys and incised 
gorges ; possible roads tend to follow contour lines along the sides of the valley 
• more important rivers (having a higher order) show broad flood plains; roads in these 
valleys appear to be piecewise linear segments that can be located at larger distances 
from the river 
In general, the order of a river segment determines the expected outline of a possible road 
as well as the distance of that road to the river. 
The hydrographic information at our disposal is in the format of a binary map (for an 
example see Fig. Hi). As the order of a river segment in its corresponding drainage network 
is crucial for the generic model, the binary map needs to be converted to a collection 
of hierarchical structures corresponding to the drainage networks. We have developed 
a classification system to produce a symbolic hierarchical description of a hydrographic 
map. The stream ordering obtained is comparable to the one proposed by Strahler [3]. For 
example, Fig. H2 shows such a representation of the map of Fig. HI. In Fig. H3 some of 
the rivers classified in this way are backprojected in map format. Each river is symbolized 
as an object having attributes to describe its end-coordinates, length, order, ... , while 
the implicit data structuring capability of the development tool is used to represent the 
drainage hierarchy. 
From the river segments derived in this way, search environments are constructed to 
perform local line extraction. A search environment is obtained by dilating the river seg 
ment as illustrated in Fig. H4 : here a SPOT image is shown and the (slightly exaggerated) 
outline of one of the segments (white) as well as the (enlarged) river-environment obtained 
from this segment (filled with the original SPOT data, here). In Figs. H5,H6 and H7 the 
contents of landcover, DTM and existing road data respectively, are shown within this 
environment. The generic model indicates that there exists a proportionality between the 
order of a river and the extent of such an environment. Given the data we are using (20 
m. grid), a factor of 3 was found to be acceptable. Additionally, the dilation procedure to 
generate river segments also has a smoothing effect on the outline of the river environment 
border : the environment will more closely approximate the flood plain for higher order 
river segments, conforming the generic model. 
In this river environment, a line detector is applied to the satellite data. Relying on 
the relation between the directional response in a given pixel and the local river direction 
perpendicular to the pixel, the magnitude response will be changed. This results in an 
operator favouring lines parallel to the river segment (see Fig. H8). To delineate a new 
road network, a tracking technique is implemented that is based on this operator as well as 
on other external data sources. To quantify constraints and relations associated with GIS 
data sources, a scoring mechanism has been designed to decide what pixel to take next 
in order to extend a line. Scores are adjusted with respect to landcover, terrain, existing 
road and magnitude and direction data. These relative score values were heuristically 
determined for a set of 5 imagery. A result of a new network derived in this way is shown 
in Fig. H9.