Tree-Component Segmentation of UAV LiDAR Point Clouds Using Attention-Enhanced PointNet++

Semantic segmentation of forest LiDAR point clouds requires distinguishing structurally similar stems, branches, and foliage-associated components. We evaluated SE-Attention, an attention-enhanced PointNet++ model with a squeeze-and-excitation (SE) channel-attention block after each of four Set Abstraction stages, against PointNet++, RandLA-Net, and KPConv on a five-class dataset. All models used XYZ coordinates, normalized relative height, and log-transformed intensity as inputs, AdamW optimization, and weighted cross-entropy combined with Lovász–Softmax. SE-Attention differed from PointNet++ only by the addition of SE modules. Across three training runs, SE-Attention achieved 89.62 ± 0.76% overall accuracy, 61.36 ± 1.74% mean intersection over union (mIoU), and 74.03 ± 1.57% mean F1-score on sampled test blocks (mean ± standard deviation). Its mean mIoU exceeded those of PointNet++, RandLA-Net, and KPConv by 13.28, 6.04, and 11.55 percentage points, respectively. Mean Stem and Branches IoU reached 50.80 ± 1.48% and 41.15 ± 2.33%, exceeding RandLA-Net by 18.79 and 14.94 percentage points. Confusion matrices averaged across the three runs showed persistent errors between woody components and Live Branches. The study provides an application-focused evaluation of established SE channel attention within PointNet++ and identifies class-specific gains and remaining errors on two plots within one forest complex; transfer to other forests was not tested.