Building detection in off-nidar very high resolution satellite images based on stereo 3D information

Abstract

Mapping or updating maps of urban areas is crucial for urban planning and management. Since buildings are the main objects in urban environments, building roof detection is an important task in urban mapping. The ideal geo-spatial data source for mapping building information is very high resolution (VHR) satellite images. On the other hand, because buildings are elevated objects, incorporating their heights in building detection can significantly improve the accuracy of the mapping. The most cost-effective source for extracting the height information is stereo VHR satellite images that can provide two types of stereo 3D information: elevation and disparity. However, most VHR images are acquired off-nadir. This acquisition type causes building leaning in the images and creates major challenges for the incorporation of building height information into roof detection. Thus, this PhD research focuses on finding solutions to mitigate the problems associated with 3D-supported building detection in off-nadir VHR satellite images. It also exploits the potential of extracting disparity information from off-nadir image pairs to support building detection. In the research, several problems associated with building leaning need to be solved, such as building roof offsetting from its footprint, object occlusion, and building façades. Moreover, the variation of the roofs offsets based on the building heights. While the offsets of building roof create difficulties in the co-registration between image and elevation data, the building façades and occlusions create challenges in automatically finding matching points in off-nadir image pairs. Furthermore, due to the variation in building-roof offsets, the mapped roofs extracted from off-nadir images cannot be directly geo-referenced to existing maps for effective information integration. In this PhD dissertation, all of the above identified problems are addressed in a progressively improving manner (i.e., solving the problems one after another while improving the efficiency) within the context of 3D-supported building detection in off-nadir VHR satellite images. Firstly, an image-elevation co-registration technique is developed that is more efficient than the currently available techniques. Secondly, the computation cost is then reduced by generating disparity information instead of the traditional elevation data. This allows bypassing a few time-consuming steps of the traditional method. Thirdly, the disparity generation is then extended from using one pair of off-nadir images to using multiple pairs for achieving an enriched disparity map. Finally, the enriched disparity maps achieved are then used to efficiently derive elevations that are directly co-registered with pixel-level accuracy to the selected reference image. Based on these disparity-based co-registered elevations, building roofs are successfully detected and accurately geo-referenced to existing maps. The outcome of this PhD research proved the possibility of using off-nadir VHR satellite images for accurate urban building detection. It significantly increases the data source scope for building detection since most (> 95%) of VHR satellite images are off-nadir and traditional methods cannot effectively handle off-nadir images.