Detecting nets of linear structures in satellite images

Fast recent progresses in satellite imagery have made available high resolution images can be analyzed with computer vision techniques. A large part of civil engineering objects appear as linear structures (e.g. channels, roads, bridges etc.). In this paper, we present a technique to detect linear features in satellite images based on an improved version of the level set extrinsic curvature. It allows the extraction of creases (ridge and valley lines) with a high degree of continuity along the center of elongated structures. However, due to its local nature, it can not cope with ambiguities originated by junctions, occlusion and branching of linear structures (e.g. hydrological or highway networks). In order to overcome this problem, we have applied a global segmentation technique based on geodesic snakes. It addresses line segmentation as a problem of detecting minimal length path. The geodesic snake looks for the path of minimal cost on a map which combines the information of the crease detector with the intensity of the original image. Some preliminary results on high resolution satellite images are presented.