SaLSE-net: a self-attention-based local structure enhancement network for landslide point cloud segmentation

Airborne Light Detection and Ranging (LiDAR) point clouds provide precise 3D ground surface data, making them valuable for automatic landslide detection. However, accurately segmenting landslide boundaries remains challenging, as existing methods struggle with boundary distinction, resulting in suboptimal overall performance. To address this gap, we propose SaLSE-Net, a self-attention-based model enhanced by local structural features for accurate landslide segmentation. SaLSE-Net operates in four steps: first, the local features of the landslide boundaries are extracted based on the variogram; second, a neighbor-aware encoder (NAEncoder) is designed to weight and aggregate neighboring points adaptively; third, it applies curvature-based downsampling to enhance landslide boundary context; and finally, the NAEncoder is embedded in each layer to integrate global and local features for precise segmentation. We applied the method to the southeastern Tibetan Plateau landslides and conducted comparative experiments. The results show that SaLSE-Net achieves a mean Intersection over Union (mIoU) of 0.791 in landslide segmentation, outperforming PointNet++ and Point Transformer models by 4.9% and 3.5%. In addition, we analyzed the optimal curvature-based downsampling threshold and the design of the neighbor-aware encoding function to achieve the best segmentation performance. Visualization of feature maps further indicates that SaLSE-Net has superior boundary prediction capability.