Morphological path filtering at the region scale for efficient and robust road network extraction from satellite imagery

Roads are important elements in geographic information systems and remote sensing applications. Their automatic extraction is challenging when only aerial or satellite images are used. Recently, some promising attempts have been made with (incomplete) path opening/closing, morphological filters able to deal with curvilinear structures. We propose here to apply morphological path filters not on pixels directly but rather on regions representing road segments, in order to improve both efficiency and robustness. The overall process is organized in two steps: first we map road segments by rectangular areas made of similar content, before we connect such segments into paths of segments or polylines using region-based path filtering. Robustness to occlusion is ensured through the adaptation of the incomplete path filtering strategy to the region scale, while better discrimination between road segments and other objects is achieved through an hit-or-miss transform that exploits background knowledge. Experiments conducted on several satellite images illustrate the interest of the proposed approach, and shows it outperforms pixelwise detection.