Evaluation of the Applicability of Automatic Segmentation Methods for Large-Scale Utilization of Airborne LiDAR Point Cloud Data

Abstract

This study evaluates the applicability of PointNet++ to airborne LiDAR point clouds based on low density and heterogeneous point distributions. Airborne point clouds in the Tama area, Tokyo were manually labeled in CloudCompare into six semantic classes (Building, Road, Vegetation, RailWay, Parking, and ElectricWire), and semantic segmentation was performed using PointNet++. Since PointNet++ relies on radius-based neighborhood search (Ball Query) originally tuned for dense indoor point clouds, directly applying the same network to airborne data leads to insufficient neighbors and unstable local feature extraction. To address this issue, we adjusted the global spatial scale of the input point clouds while keeping the network architecture and training settings unchanged, thereby ensuring a sufficient number of neighboring points within the predefined radius. The results showed that the scale factor of 0.030 yielded the most stable predictions and achieved the best mIoU, mean class accuracy (mACC), and overall accuracy (OA). These findings indicate that scale adjustment is an effective preprocessing strategy for applying PointNet++ to low-density airborne LiDAR point clouds, supporting the feasibility of semantic segmentation for wide-area environmental understanding and potential disaster-related applications.