Leveraging self-supervised point cloud pre-training for robust tree species classification in complex forest environments

This study evaluates the performance of five deep learning approaches for tree species classification using the FOR-species20K dataset, with a particular emphasis on the effectiveness of self-supervised pretraining. The experiments include supervised baseline models, such as PointNet++, DGCNN, and a standard Transformer, as well as self-supervised methods like PointContrast and Point-MAE. The supervised models were trained and tested directly on FOR-species20K, while the self-supervised models underwent pretraining on the Boreal3D dataset before being fine-tuned on FOR-species20K. The results indicate that Point-MAE significantly outperforms the supervised models, achieving higher accuracy and F1 scores, especially for species that are morphologically complex or underrepresented in the dataset. By leveraging unlabeled point clouds to learn robust representations, self-supervised methods reduce reliance on manual annotations, enhance generalization, and provide an efficient pathway for analyzing large-scale forest point clouds.