DeepLA-Net: A Power Line Point Cloud Semantic Segmentation Network Fusing Multi-Level Features and Deep Supervision

Accurate modeling and segmentation of power line point clouds are key to smart-grid inspection automation. However, existing methods struggle with the slender and highly continuous geometry of power lines, the sparsity of point clouds, and complex background clutter. This paper presents DeepLA-Net, a framework for identifying and segmenting key infrastructure (e.g., power lines and transmission towers) from Uninhabited aerial vehicles (UAVs) equipped with light detection and ranging (LiDAR) point clouds. Built on the Point Transformer V3 (PTv3) backbone, the approach employs a residual local feature extraction (ResLFE) module and a self-attention aggregation (SAA) module, together with a hybrid deep supervision (HDS) strategy. ResLFE preserves fine-grained geometry via residual connections in shallow layers. SAA models long-range dependencies through a learnable gating scheme to sustain topological coherence. HDS injects auxiliary supervision into multiple decoder layers to accelerate convergence, improve training stability, and enhance generalization. Comprehensive experiments on a self-collected UAV LiDAR dataset demonstrate the effectiveness of each component through ablations and show that DeepLA-Net attains superior segmentation of power lines and towers compared with PointNet, PointNet++, PTv2, and PTv3.