An efficient approach for image-DSM co-registration for urban building extraction

Very high resolution (VHR) satellite images are the ideal geo-data for mapping urban areas. The co-registration of such images with elevation data is crucial for accurate 3D-supported building extraction and mapping applications. However, VHR satellite images are usually acquired off-nadir. Over urban areas, off-nadir images suffer from severe building lean caused by the images' perspective view. On the other hand, the elevations of digital surface models (DSM) are usually of orthographic projection. Such a difference makes pixel-by-pixel co-registration very challenging unless the DSM data are modified to be of Line-of-Sight projection (LoS-DSM). Therefore, this paper introduces a novel image-DSM co-registration method for building extraction. Based on generating disparity maps, the method constructs a perfectly co-registered LoS-DSM which is more efficiently than traditional algorithms. The root-mean-square-error of the developed LoS-DSM elevations was found to be less than 2 pixels relative to the traditional photogrammetric approach. Additionally, these elevations are of pixel-level co-registration accuracy.