Airborne LiDAR point-cloud semantic segmentation is fundamental to three-dimensional (3D) scene understanding and supports applications in smart cities, urban modeling, infrastructure extraction, land surveying, and environmental monitoring. Existing methods often inadequately coordinate local geometric representation with contextual modeling. To address this limitation, we propose LSENet (local subdomain-aware and edge-enhanced dynamic fusion network). LSENet extracts fine-grained geometric features through adaptive local-subdomain partitioning, facilitates information exchange through cross-subdomain interactions, and improves feature representation and boundary discrimination through an edge-enhanced dual-attention fusion module. Experiments on the Vaihingen 3D and DFC2019 datasets yielded Overall Accuracies (OAs) of 83.9% and 97.8% and mean F1 scores of 70.5% and 89.1%, respectively, demonstrating the effectiveness of LSENet and its applicability across scenes.LSENet uses a U-Net-style encoder–decoder architecture comprising three core modules. The Local Subdomain Space-aware feature Encoding and Aggregation (LSSEA) module constructs an adaptive local coordinate system using principal component analysis (PCA) and encodes relative position, spatial distribution, orientation, and curvature to obtain fine-grained geometric representations. The Local Subdomain feature Enhancement and Interaction Attention (LSEIA) module enhances features within each subdomain and establishes cross-subdomain interactions to improve local contextual modeling. The Edge-enhanced Grouped Multi-scale Dual-Attention dynamic fusion (EGMDA) module dynamically fuses edge-enhanced grouped channel attention with multi-scale spatial attention, thereby strengthening boundary representation and classification accuracy.Comparative experiments were conducted 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%, yielding the highest OA among the compared methods. Performance improved substantially for challenging classes, including power lines, roofs, and trees, confirming the effectiveness of the proposed modules in preserving complex structures and distinct boundaries. On DFC2019, LSENet achieved an OA of 97.8% and a mean F1 score of 89.1%; it again ranked first in OA and showed strong cross-scene generalizability. Visual comparisons further indicate that LSENet produces more spatially coherent regions and clearer boundaries than the baseline.LSENet integrates local geometric awareness, cross-subdomain interaction, edge-preserving dual attention, and dynamic fusion to address limitations of existing point-cloud segmentation networks. Results on two airborne LiDAR benchmarks confirm its effectiveness, robustness, and cross-scene generalizability. Future work will develop more discriminative geometric descriptors and lightweight dynamic-inference strategies and will evaluate LSENet on newly released, large-scale, diverse airborne LiDAR datasets to improve its applicability in real-world settings.