AUTOMATIC ROAD EXTRACTION FROM MULTISPECTRAL HIGH RESOLUTION SATELLITE IMAGES

In this paper we propose an approach for automatic road extraction from high resolution multispectral imagery, such as IKONOS or Quickbird, in rural areas. While aerial imagery usually consists of 3 spectral bands, high resolution satellite data comprises 4 spectral bands with a better radiometric quality compared to film, but a worse geometric resolution. Therefore, strongly making use of the spectral properties of satellite imagery is a way to mitigate the geometric disadvantages and achieve results comparable to those from aerial imagery. To this end, we employ local as well as global properties of roads. The extraction starts with the extraction of Steger lines in all spectral channels. The lines are used as cues for roads to generate training areas for a subsequent automatic supervised classification. The result of the classification, the road class image with its well behaved characteristics, is used as an additional source for the extraction of road candidates. Our novel verification process for road hypotheses makes use of geometric conditions as well as the spectral properties of roads by computing the road energy from the road class image. From the verified road hypotheses a final road network is generated by first bridging small gaps based on a weighted graph and then searching for missing connections in the network by calculating local detour factors. The missing connections are closed by optimizing ziplock snakes between pairs of seed points and are then verified. An evaluation of the results is carried out by comparing our results with manually extracted reference data demonstrating the potential as well as the problems of the approach.