From Unmanned Aerial Vehicle–based Laser Scanning to Airborne Laser Scanning: Direct Inference and Fine-tuning of Geometry-only Deep Learning Models for Individual Tree Species Classification

Tree species classification from lidar point clouds is critical for forest monitoring and large-scale ecological assessments. Deep learning models have shown strong performance when trained and tested on data collected from the same platform, but it remains unclear whether a model trained on unmanned aerial vehicle–based laser scanning (ULS) point clouds can be directly applied to other data types, such as airborne laser scanning (ALS), without additional training. This study investigates geometry-only tree species classification from cross-platform lidar point clouds and evaluates how ULS-trained models perform on ALS data through direct inference and fine-tuning. Four deep learning models, PointNet, PointNet++, Point Transformer V3 (PTv3), and three-dimensional dual-attention glocal transformer network (3DGTN), were evaluated at four point sampling sizes (256, 512, 1024, and 2048 points per tree), together with original point-count statistics and a sampling-strategy sensitivity analysis. All four models achieved stable within-platform ULS performance, with overall accuracy (OA) ranging from approximately 75.0% to 80.5%. Direct ULS-to-ALS inference was strongly model-dependent: 3DGTN, PointNet++, and PTv3 transferred poorly, whereas PointNet achieved higher OA but lower macro-level metrics, indicating persistent species-level imbalance and confusion. Fine-tuning with ALS data improved cross-platform performance across models, with the best OA reaching 80.7% for PointNet and 77.1%, 76.7%, and 76.4% for 3DGTN, PointNet++, and PTv3, respectively. These results show that cross-platform tree species classification is sensitive to the combined differences in point density, scanning geometry, and structural completeness and that fine-tuning with target-domain ALS data can reduce this domain gap but cannot resolve species imbalance or platform-specific structural differences. The source code used in this study is publicly available at https://github.com/ssamy001/Tree-species-classification