Community detection framework based on 3D shape descriptors for tree species classification in point cloud data

Accurate tree species classification from remote sensing data, such as LiDAR point clouds, is important for various applications, including vegetation monitoring and forest growth prediction. Although numerous machine learning algorithms are used widely for these tasks, several challenges remain. These include the need for extensive training datasets, over-fitting to specific geographic areas and environmental conditions during tree growth, the requirement for post-processing adjustments, and unreliable performance with rare tree species and shapes. This paper presents a robust framework for classifying tree species directly from diverse point cloud datasets, eliminating the need for training machine learning models or manually preparing the training datasets. The proposed approach performs community detection on a graph using features extracted from point clouds of individual trees. Multiple shape descriptors are proposed as feature vectors that consider 3D tree crown structure and are rotationally invariant. Community detection groups the trees into distinct communities based on these feature vectors. Tree species classification is performed by classifying these communities, reducing the manual effort significantly, as only a few trees from each community need inspection. The proposed framework was validated using both real-world terrestrial LiDAR and synthetic point clouds. The results demonstrate its competitiveness with established methods, while surpassing the performance of traditional clustering techniques applied to the same feature vectors. Additionally, the results confirm the effectiveness of the proposed feature vectors for achieving competitive tree species classification accuracy.