Adaptive Shortest Path Tracking for Robust Leaf–Wood Separation in Individual Trees From TLS Point Clouds

Leaf-wood separation is crucial for single-tree aboveground biomass estimation and three-dimensional reconstruction. Although the nondestructive and efficient acquisition of fine-grained, high-density point cloud data can be performed using terrestrial laser scanning technology, existing methods suffer from various drawbacks, including insufficient detection of fine branches, limited robustness to point cloud subsampling, and weak adaptability across different tree species and crown structures. A core issue lies in the over-reliance on prior values for key algorithm parameters. This article proposes an adaptive shortest path tracking for robust leaf–wood separation (ASPTS) in individual trees. First, a graph is constructed, and the shortest path backtracking is employed to extract skeleton points. Second, an improved k-nearest neighbor algorithm is proposed to adaptively optimize the number of neighboring points based on the shortest path, thereby obtaining initial wood points. Third, the feature descriptor construction for characterizing trunk and branch structures is optimized using principal component analysis by implementing an enhanced adaptive neighborhood radius selection strategy. Finally, final wood points are extracted using a region-growing approach guided by a stepwise feature thresholding scheme. A total of 22 individual trees, which represent different species, heights, and crown structures, are selected as test subjects. The results demonstrate the capability of ASPTS to make a good balance between type I and type II errors. ASPTS consistently exhibits strong fine-branch detection capability and robust performance under varying conditions, including different tree species, crown structures, and point cloud densities. ASPTS demonstrates superior performance compared to four state-of-the-art methods.