Spatial Feature Structure Generation Network for Airborne LiDAR Point Cloud Semantic Segmentation

With the rapid development of airborne Light Detection and Ranging (LiDAR) sensors, point clouds have been widely applied in fields such as surveying and mapping engineering, intelligent monitoring, and autonomous driving. However, the irregular structure of point clouds seriously affects the segmentation, recognition, and understanding of point clouds. To address the challenge, this paper proposes a spatial feature structure generation network (SFSGNet). The spatial feature structure generation (SFSG) module can deeply fuse attribute features and spatial coordinate features to generate highly expressive spatial structure features in both semantic and spatial contexts, and through the spatial structure features, SFSGNet can effectively handle irregular point clouds, describe the structure of point cloud objects, and achieve accurate point cloud semantic segmentation. We evaluate our SFSGNet using the ISPRS 3D Labeling Benchmark dataset, DALES dataset and OpenGF dataset. Experimental results show that our SFSGNet achieves an average F1 score 74.4% and overall accuracy 85.4% on the ISPRS dataset, achieves an average F1 score 89.3% and overall accuracy 98.1% on the DALES dataset, and achieves an average F1 score 98.2% and overall accuracy 98.5% on the OpenGF dataset. Compared with other representative models, our SFSGNet demonstrates excellent segmentation performance and strong generalization.