DTM Extraction and building detection in DSMs having large holes

Digital Surface Models (DSMs) and Digital Terrain Models (DTMs) are crucial for automatic detection of buildings and other ground features in urban areas. They are the basis for numerous geospatial applications, such as 3D city modelling, 3D mapping infrastructures, drainage pattern studies, etc. LiDAR based DSMs usually contain less holes than image-based DSMs. This is mainly because of the mismatching error acquired due to occlusion in photogrammetry. However, if high-rise buildings are present in the scene, large holes are most likely to occur in LiDAR derived DSM. In this research, a modified algorithm for LiDAR-based DSM to DTM filtering approach is presented. The algorithm deals with the original DSM without aiming for precise outcomes. Firstly, pixels representing holes are automatically detected and labelled with unique value. Since available DTM extraction algorithms are mainly based on finding pixels having minimal elevation, this unique value must belong to the off-terrain to avoid using them as minima. Afterwards, the Network of Ground Points algorithm is applied for extracting DTM from DSM to be further used for building detection purposes. The proposed method is tested on the benchmark dataset for Toronto city provided by the ISPRS working group III / 4. The evaluation analysis was applied based on calculating correctness and completeness of the detected building segments. Results showed the practical effectiveness of the proposed technique in achieving average correctness of 91 %.