ALS Point Cloud Semantic Segmentation Based on Graph Convolution and Transformer With Elevation Attention

Semantic segmentation of airborne point clouds is crucial for 3D scene reconstruction and remote sensing in surveying applications. Current deep learning methods for point clouds primarily focus on effectively aggregating local neighborhood information. However, they often overlook the fusion of global context information and elevation features, which are vital for airborne point clouds. In this study, we propose Dense-LGEANet, a novel network with dense connected architecture and multiscale feature supervision based on our designed LGEA module. The key component of our LGEA module is the combination of the graph convolution block and the transformer block with elevation attention. It can effectively fuse local neighborhood information and global context information to improve the accuracy of semantic segmentation of airborne point cloud. Moreover, the designed dense connected network architecture can enhance the feature extraction capability for point cloud objects at different scales by facilitating interactions between multiple up-sampling and down-sampling layers. We have conducted multiple experiments on the public point cloud dataset. Experimental results show that our method can achieve an mIoU of 58.5% and an mF1 of 72.0% on the ISPRS Vaihingen 3D dataset, while an mIoU of 67.2% and an mF1 of 78.3% on the LASDU dataset. It demonstrates the superior performance of our network and the effectiveness of the proposed feature enhancement module and network architecture.