Tree Species Classification from TLS Point Clouds Using Multi-Task Learning and Woody-Only Point Cloud Generation

Terrestrial Laser Scanning (TLS) can provide detailed three-dimensional structural information for individual trees and has become an important data source for tree species classification. However, most existing models are trained using leaf-on point clouds and therefore tend to rely heavily on leaf distribution and crown appearance. When the input changes from leaf-on point clouds to woody-dominated representations, classification performance often declines. To address this issue, this study proposes a mixed-input tree species classification framework for six typical temperate broadleaf tree species. First, a KPConv-based wood–leaf separation model was used to extract woody point sets from leaf-on TLS point clouds, thereby generating woody-only representations for subsequent classification. Second, a multi-task learning network based on DGCNN was constructed. In addition to the main task of tree species classification, an auxiliary task for input-representation discrimination was introduced to enhance the model’s adaptability to different input forms. Experiments were conducted using a dataset composed of local TLS samples from China and publicly available single-tree point clouds from the BioDiv dataset. The results show that the proposed method achieved an overall accuracy of 94.3% on the mixed test set of six typical broadleaf tree species, with average Precision, Recall, and F1 values of 94.3%, 93.6%, and 93.9%, respectively. These results indicate that integrating woody structural representations with multi-task learning can effectively alleviate overreliance on leaf-on appearance features and improve classification robustness under different input representations.