Abstract: [Purposes] To address the challenges of large spans in transmission corridors,strong vegetation and terrain obstructions,and the high cost of acquiring finely annotated samples,this paper proposes a point cloud-image fusion semi-supervised segmentation and rapid vectorization method tailored for corridor risk components.[Methods] A dual-branch segmentation model integrating point cloud and imagery data was constructed.The imagery branch extracts multi-scale texture semantics,while the point cloud/elevation branch characterizes geometric and spatial relationships.Cross-modal attention achieves feature alignment and complementarity.Utilizing the spatial intersection of canopy height models and corridor buffers to automatically generate binary risk grids as weak annotations,we introduced teacher-student consistency learning and a confidence-based pseudo-labeling mechanism.Combined with Focal and Lovsz losses,this mitigates class imbalance and enhances segmentation accuracy for slender objects.During inference,connected component analysis and area threshold filtering are applied to segmentation results,overlaid with corridor boundary constraints to generate structured vector alert features.[Findings] Experiments on actual transmission line data in Guangdong Province demonstrate that this method outperforms unimodal approaches relying solely on point cloud or optical image features across three metrics： mIoU, F1 score,and overall accuracy.Specifically,it achieves mIoU of 0.185,F1 score of 0.307,and overall accuracy of 0.842,while generating topologically complete corridor risk patches.[Conclusions] The proposed point cloud-image fusion semi-supervised segmentation and rapid vectorization method enables effective extraction of high-vegetation/potential over-line objects within transmission corridors and their vectorization for operation and maintenance purposes,with minimal additional manual annotation costs.This approach demonstrates engineering feasibility for practical applications．

Key words: transmission corridor, risk component identification, semi-supervised semantic segmentation, weakly supervised annotation generation, cross-modal feature alignment, vectorization of risk targets