Leveraging Domain Characteristics to Refine Deep-Learning-Based Semantic Segmentation of Outdoor Point Clouds

High-resolution point clouds generated by modern LiDAR sensors on drones enable point clouds of much higher resolution and detail, presenting new challenges for semantic segmentation, such as efficiency, limited receptive fields, and implausible class prediction results. To address these, we integrate relative elevation and differential morphological profiles - both of which are domain-specific features in remote sensing - into RandLA-Net for 3-D point cloud segmentation. This enables enhanced feature representation without increasing network complexity or requiring input downsampling. On the Hessigheim dataset, including relative elevation and morphological profiles improves the m F1 score by +5.52%. In addition, we utilize conditional random fields (CRFs) with an interclass reliability matrix to refine predictions and enforce realistic class neighborhoods, further increasing the m F1 score to exceed 78%. Overall, this approach ensures accurate and efficient segmentation, leveraging domain-specific preprocessing characteristics and domain knowledge about class neighborhoods. A comparison with competing methods, mostly favoring our approach, indicates that all deep-learning networks operating on remote sensing point clouds could benefit from explicit incorporation of these domain characteristics.