Assessing Applications of Self-Supervised Learning for Tree Species Classification from LiDAR Point Clouds

Abstract. Individual tree species classification from LiDAR (Light Detection And Ranging) point clouds has significant potential to support forest inventory and management, yet remains challenging due to complex three-dimensional canopy structures and the limited availability of labelled ground truth data. This study investigates self-supervised learning for tree species classification from LiDAR point clouds by comparing the PointMAE, a masked autoencoder-based model, with two supervised baselines, PointNet and PointNet++. Using the FOR-species20k dataset, two experiments were conducted: a 33-species classification and a 6-species classification, each evaluated with point cloud sizes of 2048 and 8192 points. Using 2048 points, the PointMAE achieved the highest overall accuracy in both experiments (0.67 and 0.89 respectively), outperforming PointNet++ (0.63 and 0.84) and PointNet (0.39 and 0.75). Across all models, performance decreased when using 8192 points, indicating sensitivity to point cloud density and sampling. Per-species analysis showed that coniferous species with distinctive crown geometries were the easiest to classify, while broadleaf species with similar crown forms, particularly Carpinus betulus, were the most challenging. These results show that self-supervised pretraining can improve classification accuracy over fully supervised approaches, highlighting its value for forestry applications where labelled data are limited.