International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
  
  
Figure 6: (Left) Line strength image representing the absolute 
values of the maximum negative eigenvalues of the Hessian ma- 
trix. (Right) Gradient image representing the absolute values of 
the sobel filter mask. The original image is given in figure 2. 
as well as the mixed derivative. Large negative eigenvalues are 
obtained for a bright line, whereas large positive eigenvalues cor- 
respond to dark lines. We note that also point features lead to 
large positive as well as negative eigenvalues, but they do not dis- 
turb our procedure. As experience shows, that most of the roads 
appear as bright lines in the images, only the maximum negative 
eigenvalues are used. To take into account the information of all 
image channels, the eigenvalues are calculated for every image 
channel and the maximum of the absolute value for every pixel 
is written into the line strength image. An example is shown in 
figure 6, left. 
Also the maximum gradient image is generated using all image 
channels. For every image channel a Sobel image is calculated 
using the medium absolute values. The maximum value for every 
pixel of these Sobel images is taken for the maximum gradient 
image (cf. fig. 6, right). By employing the line strength and the 
gradient image as input data for our snake based approach we 
focus on the features we are interested in, i.e., bar-shaped roads 
and field borders. 
3.4 Verification of Road Sections 
A disadvantage of snakes is that they will minimize their energy 
in any case, even if there are no meaningful image features avail- 
able. This leads to the necessity to verify the result by examining 
the path of the resulting snake. In our approach the verification 
of the snake is synonymous with the verification of the road sec- 
tions. As criteria for the correctness the line or edge strength 
along the path of the snake is used. A section is verified only if 
there is enough evidence for a linear feature along the path. 
A grey value profile in the line strength or gradient image per- 
pendicular to the snake direction is calculated for every snake 
point. To evaluate the quality of a single point, the profile is 
first smoothed with a Gaussian kernel. Then, the maximum value 
along the profile and the position of the maximum are calculated. 
For a valid point the maximum should be close to the center of 
the profile and the second derivative along the profile at the max- 
imum point should be significantly smaller than zero. To accept 
a road section, the percentage of valid points needs to be larger 
than a given threshold. 
3.5 Global Grouping 
The result of the previous step are individual road sections. To 
construct a network, road sections are grouped into larger struc- 
1058 
tures. As road sections may result from either line or edge fea- 
tures they need to be fused first into linear features. To do so, 
the road sections are evaluated using the quality measure gener- 
ated by the previous verification. Because road sections resulting 
from lines are more reliable, they are given a higher weight than 
those obtained from edges. The result are the road sections with 
the highest evaluation value. 
An important property of a road network is, that most points on 
the network can be reached from all other points along an optimal 
path with a minimum detour. To make use of this property, we 
generate link hypotheses according to (Wiedemann and Ebner, 
2000). The distance between pairs of points within the network 
is calculated along the existing network and along a hypothetical 
optimal path, for which the Euclidean distance is used. To form 
so-called preliminary link hypotheses, a detour factor is calcu- 
lated as follows: 
; network distance 
detour fadtor= ———>x""" (3) 
optimal distance 
The link hypotheses are checked starting with the hypotheses 
with the largest detour factor. If a link hypotheses is accepted, 
the new connection is inserted into the road network. Due to 
changes in the network, the generation of link hypotheses has to 
be repeated. This is iterated until no more new link hypotheses 
are generated. The result of the global grouping is the final road 
network. 
4 EXPERIMENTAL RESULTS 
For the validation of the proposed approach pan-sharpened IRS- 
1C/D satellite images for a test site in northern Africa were used. 
The images were selected in a way that they comprise different 
road types in agricultural as well as in mountainous test arcas. 
We have not yet generated reference data for a quantitative evalu- 
ation, therefore, the validation is done just qualitatively by visual 
inspection. 
Figure 7 shows the results obtained for the first image sample 
shown in figure 2. The network extracted for this example shows, 
that in agricultural areas not only main roads can be extracted, 
but also smaller roads that connect, e.g., individual farms, for 
instance on the right side of figure 7. One difficulty here is the 
distinction between roads and paths that follow the borders of a 
field. 
The second example (fig. 8 and 9) shows a complex road net- 
work. Several small villages are connected by roads of different 
importance. Most of the roads outside the villages were correctly 
extracted. As the approach is developed for roads in agricultural 
areas, the streets inside the villages could, as was to be expected, 
mostly not be extracted. Worth to mention is the small number of 
false positives. 
The third example (fig. 10 and 11) shows a road passing through a 
mountainous area. For this example the approach of (Wiedemann 
et al., 1998) was used. No limitation for the maximal allowed 
curvature was set.