The search window is considered as an image and, to perform F* algorithm, we must define a cost for each pixel of the search window. The probability for the pixel to be on a linear feature induces its cost value. We have to consider the maximum of characteristics of linear features to define this cost.

The cost value will be a weighted mean of all the expressions corresponding to all the characteristics.

1.

The first characteristic is the homogeneity of radiometry along the linear feature on SPOT image.

The only point of the searched linear feature that we know at the beginning is the starting point. So, we initialize the radiometry mean of the element that we search, with the radiometry of the starting point and recalculate it after each iteration with all the pixels which are on the resulting path. A good expression for the pixel cost can be the absolute value of the difference between the radiometry mean and the pixel radiometry. SPOT P and XS3 are the best bands for this characteristic computation as linear features are homogeneous and there is a contrast with environment.

2.

The second characteristic is based on detection of ``roof'' or ``valley profile'' edges. We can compute a 2nd derivative filter (ISEF Roof and Valley filter) on the SPOT image(s) and obtain, with a disjunctive fusion, a good image with high values ''V'' on linear features. If the image has 256 levels of grey, (255-V) is an expression for the pixel cost.

                  
  
 IRIT-UPS