Large-Scale DTM Generation From Satellite Data

Liuyun Duan, Mathieu Desbrun, Anne Giraud, Frederic Trastour, Lionel Laurore; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2019, pp. 0-0

In remote sensing, Digital Terrain Models (DTM) generation is a long-standing problem involving bare-terrain extraction and surface reconstruction to estimate a DTM from a Digital Surface Model (DSM). Most existing methods (including commercial software packages) have difficulty handling large-scale satellite data of inhomogeneous quality and resolution, and often need an expert-driven manual parameter-tuning process for each geographical type of DSM. In this paper we propose an automated and versatile DTM generation method from satellite data that is perfectly suited to large-scale applications. A novel set of feature descriptors based on multiscale morphological analysis are first computed to extract reliable bare-terrain elevations from DSMs. This terrain extraction algorithm is robust to noise and adapts well to local reliefs in both flat and highly mountainous areas. Then, we reconstruct the final DTM mesh using relative coordinates with respect to the sparse elevations previously detected, and induce preservation of geometric details by adapting these coordinates based on local relief attributes. Experiments on worldwide DSMs show the potential of our approach for large-scale DTM generation without parameter tuning. Our system is flexible as well, as it allows for a straightforward integration of multiple external masks (e.g., forest, road line, buildings, lake, etc) to better handle complex cases, resulting in further improvements of the quality of the output DTM.