International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
  
  
Figure 1: Sample images from mountainous (left) and desert ar- 
eas (right) 
2 MODEL AND STRATEGY 
Due to large differences in the appearance of roads in different 
areas in north Africa a single model for automatic road extraction 
is insufficient. We distinguish three areas: agricultural, moun- 
tainous, and desert. The characteristics of roads in IRS satellite 
images in these areas can be described as follows (cf. fig. 1 and 
2): 
In mountainous areas roads are strongly affected by the to- 
pography. Roads often turn with a large curvature or even with 
sharp bends. In the images the roads are mostly represented as 
bright and only seldom as dark lines. 
In desert areas roads mostly appear as bright or dark lines with 
few disturbing objects. The distinction from other linear objects, 
e.g., pipelines, is often difficult. 
In agricultural areas roads appear as elongated structures. They 
often have no bar-shaped line profile in the images, but can be 
seen indirectly as collinear edges of field borders. 
A distinction to what type of area a region belongs can be done 
mostly automatically based on a Digital Terrain Model (DTM) 
and the image data itself. Agricultural areas show high intensities 
in the near infrared channel, mountainous areas are characterized 
by extended steep slopes in the DTM, and desert areas consist 
of homogeneous surfaces with low intensities in the near infrared. 
Road extraction in mountainous and desert areas starts by 
extracting lines with the Steger extractor (Steger, 1998). All 
spectral channels are used independently. The resulting sub-pixel 
lines are evaluated and fused. In mountainous areas there is 
no limitation in curvature, whereas in desert areas only linear 
features with a small curvature are accepted. The verified and 
fused lines are globally grouped into the road network. A 
detailed description of the approach is given in (Wiedemann et 
al., 1998). 
The extraction of roads in agricultural areas is much more chal- 
lenging than in the other two areas. Here, roads not always appear 
as lines (cf. fig. 2) because they run in many cases along field 
borders. This means that the borders of the fields often indirectly 
represent the path of the road. On the other hand, these roads usu- 
ally form elongated collinear or curvilinear structures with small 
1056 
  
Figure 2: Sample image from an agricultural area. The white 
rectangle shows the part used in figures 4 and 6. 
curvature, i.e., they can be approximated by linear structures. 
Borders of fields means that there is a more or less strong grey 
value gradient perpendicular to the road direction. The proposed 
approach uses these characteristics to construct road sections. To 
employ as much information as possible, both lines and edges are 
used to form possible road connections. A detailed description of 
the extraction is given in the next section. 
3 ROAD EXTRACTION AS COLLINEAR FEATURES 
Our goal is to group roads appearing as lines and edges of the 
field borders into longer linear structures and by this means rec- 
ognize and delineate the roads. We start with the extraction of 
lines and edges (cf. fig. 3 and 5), both termed linear features 
for the remainder of this paper. From these features connection 
hypotheses are constructed and evaluated. The best path for the 
connection is obtained by optimizing a ziplock snake between 
the two adjacent endpoints of the linear features. The final road 
network is obtained by globally grouping the road sections. 
3.1 Extraction of Linear Features 
The extraction of the linear features is performed with the Steger 
sub-pixel line- and edge extractor. The extracted features are split 
into segments with a curvature below a given threshold. This is 
done for all image channels independently. In a following step 
the resulting lines and edges from all image channels are fused to 
single data sets. From these data sets connections are constructed. 
Figure 4 shows results of line and edge extraction. 
3.2 Construction of Connections 
Connections consist of elongated features with a small curvature. 
Two linear features are used to construct a connection if they sat- 
isfy the following conditions: 
e the linear features have to be collinear (jc, ,) 
e the linear feature and their straight connection must be 
collinear (fic) 
    
  
   
  
  
  
    
   
     
   
    
    
     
     
   
    
   
  
  
    
    
  
   
  
  
   
    
   
   
   
    
   
  
  
    
  
Interr 
  
Figur 
image 
Figure 
optimi