机载LiDAR点云语义分割是三维地物精细化解译与场景理解的重要基础，在智慧城市建设、城市三维建模、交通基础设施提取、国土空间调查和生态环境监测等领域具有重要的应用价值。针对现有深度学习点云语义分割方法在局部几何结构刻画与上下文建模之间协同不足的问题，提出了一种新的机载LiDAR点云语义分割网络LSENet（Local Subdomain-aware and Edge-enhance Dynamic Fusion Network）。首先通过局部子域自适应划分提取每个采样点细粒度空间几何特征；其次在子域间引入特征增强的交互机制以强化局部空间感知与信息流通；最后设计了边缘增强的分组多尺度双注意力动态融合模块，实现了关键特征的动态加权融合，增强了不同语义类别之间的可分性，并提升了模型对复杂结构及边界区域的判别能力。基于Vaihingen 3D与DFC2019两个公开数据集的实验结果表明，LSENet分别取得了83.9%和97.8%的总体精度、70.5%和89.1%的平均F1分数，整体性能优于部分对比方法，验证了所构建网络在机载LiDAR点云语义分割中的有效性和跨场景适用性。

Abstract: [Objective] Airborne LiDAR point cloud semantic segmentation is fundamental to 3D sce-ne understanding and has important applications in smart cities, urban modeling, infrastructure ex-traction, land surveying, and environmental monitoring. To address the insufficient coordination be-tween local geometric representation and contextual modeling in existing methods, we propose LSE-Net (Local Subdomain-aware and Edge-enhanced Dynamic Fusion Network). LSENet extracts fi-ne-grained geometric features through adaptive local subdomain partitioning, enhances information exchange via cross-subdomain interactions, and improves feature representation and boundary dis-crimination through an edge-enhanced dual-attention fusion module. Experiments on the Vaihingen 3D and DFC2019 datasets achieve Overall Accuracy (OA) values of 83.9% and 97.8%, and mean F1-scores of 70.5% and 89.1%, demonstrating its effectiveness and cross-scene applicability. [Meth-od] LSENet adopts a U-Net-style encoder–decoder architecture with three core modules. The Local Subdomain Space-aware feature Encoding Aggregation module (LSSEA) constructs an adaptive lo-cal coordinate system using PCA and encodes relative position, spatial distribution, orientation, and curvature information for fine-grained geometric representation. The Local Subdomain feature En-hancement and Interaction Attention module (LSEIA) enhances intra-subdomain features and es-tablishes cross-subdomain interactions to improve local contextual modeling. The Edge-enhanced Grouped Multi-scale Dual-Attention dynamic fusion module (EGMDA) integrates edge-enhanced grouped channel attention and multi-scale spatial attention through dynamic fusion, strengthening boundary representation and classification accuracy. [Result] Comparative experiments were con-ducted on two widely used benchmark datasets, ISPRS Vaihingen 3D and DFC2019. On Vaihingen 3D, LSENet achieved an OA of 83.9% and a mean F1-score of 70.5%, obtaining the best OA among compared methods. Significant improvements were observed in challenging categories such as power lines, roofs, and trees, demonstrating the effectiveness of the proposed modules in preserving com-plex structures and clear boundaries. On DFC2019, LSENet achieved an OA of 97.8% and a mean F1-score of 89.1%, ranking first in OA and demonstrating strong cross-scene generalization. Visuali-zation results further show that LSENet produces more coherent regions and clearer boundaries than the baseline. [Conclusion] LSENet effectively integrates local geometric awareness, cross-subdomain interaction, and edge-preserving dual attention with dynamic fusion to overcome limitations of ex-isting point cloud segmentation networks. Results on two airborne LiDAR benchmarks validate its effectiveness, robustness, and generalization capability. Future work will focus on developing more discriminative geometric descriptors, lightweight dynamic inference strategies, and evaluating the proposed method on newly released large-scale and diverse airborne LiDAR datasets to improve its applicability in real-world scenarios.