Registration of satellite images based on road network map

Abstract

This paper presents a novel approach to fully automatic satellite image registration to a local earth coordinate system based on the digital map of the road network. Automatic satellite image co-registration methods typically establish correspondences between points on the target and registered image using local image features such as SIFT, however they are not effective for multi-temporal and multi-sensor registration. As the road network is a prominent feature on the satellite images and is visible in most circumstances, we propose to map it directly to the digital road network topographic map, which is readily available in GIS databases. Our approach starts by detection of road segments in a satellite image, producing a road mask image. Small road mask patches are matched against rasterized road network map to find a set of probable patch locations. Following optimization of translation parameters, each satellite image tile is coarsely located on a road map. In a second fine optimization step, smaller road patches locations are further optimized. The algorithm produces a set of tie points, which include coordinates of satellite image pixels within local metric coordinate system. Experiments were performed on a set of 5 RapidEye images of Slovenia from different periods of the year. The results show that the average error of the points is below 1 pixel for all images, and that the reliability in terms of outlier pixels is very high.