A Novel Supervoxel-Based NE-PC Model for Separating Wood and Leaf Components from Terrestrial Laser Scanning Data

The precise extractions of tree components such as wood (i.e., trunk and branches) and leaves are fundamental prerequisites for obtaining the key attributes of trees, which will provide significant benefits for ecological and physiological studies and forest applications. Terrestrial laser scanning technology offers an efficient means for acquiring three-dimensional information on tree attributes, and has marked potential for extracting the detailed tree attributes of tree components. However, previous studies on wood–leaf separation exhibited limitations in unsupervised adaptability and robustness to complex tree architectures, while demonstrating inadequate performance in fine branch detection. This study proposes a novel unsupervised model (NE-PC) that synergizes geometric features with graph-based path analysis to achieve accurate wood–leaf classification without training samples or empirical parameter tuning. First, the boundary-preserved supervoxel segmentation (BPSS) algorithm was adapted to generate supervoxels for calculating geometric features and representative points for constructing the undirected graph. Second, a node expansion (NE) approach was proposed, with nodes with similar curvature and verticality expanded into wood nodes to avoid the omission of trunk points in path frequency detection. Third, a path concatenation (PC) approach was developed, which involves detecting salient features of nodes along the same path to improve the detection of tiny branches that are often deficient during path retracing. Tested on multi-station TLS point clouds from trees with complex leaf–branch architectures, the NE-PC model achieved a 94.1% mean accuracy and a 86.7% kappa coefficient, outperforming renowned TLSeparation and LeWos (ΔOA = 2.0–29.7%, Δkappa = 6.2–53.5%). Moreover, the NE-PC model was verified in two other study areas (Plot B, Plot C), which exhibited more complex and divergent branch structure types. It achieved classification accuracies exceeding 90% (Plot B: 92.8 ± 2.3%; Plot C: 94.4 ± 0.7%) along with average kappa coefficients above 80% (Plot B: 81.3 ± 4.2%; Plot C: 81.8 ± 3.2%), demonstrating robust performance across various tree structural complexities.