Dynamic region growing approach for leaf-wood separation of individual trees based on geometric features and growing patterns

The separation of leaf and wood points remains challenging due to the diversity of tree species and structures. We propose an automatic leaf-wood separation method from tree point clouds, leveraging trees' shape information, growing patterns, and topological information for higher separation accuracy. First, the tree point clouds are roughly classified into initial planar, linear, and scattered points based on the dimensional structure of each point. Then, a slicing strategy combined with the DBSCAN (density-based spatial clustering of application with noise) clustering algorithm is used to exploit the potential wood clusters in each layer. A novel dynamic region growing algorithm based on an energy function is proposed for extracting tree trunk and large branches from initial planar points, incorporating the geometric features and growing patterns of trees. Finally, a novel inner-to-outer region growing method is proposed to extract the tiny branches based on voxel density and the principal direction. Twelve tree samples covering twelve species and locating tropical, temperate, and boreal areas are used to evaluate the method's performance. The mean separation accuracy and kappa coefficient of the proposed method are 0.942 and 0.805, respectively. Experimental results indicate that the proposed method exhibits a strong ability to distinguish the wood and leaf points for various trees with different species and structures.