Tree architecture, characterized by the three-dimensional (3D) arrangement of branches, plays a critical role in regulating key ecological functions such as light interception, resource transport and structural stability. Terrestrial laser scanning (TLS) has emerged as a powerful tool for capturing detailed and accurate structural information of trees in complex forest environments, making it a promising technique for quantifying tree architecture. However, effective wood–leaf separation—a critical prerequisite for reconstructing three-dimensional tree models from TLS data—remains a significant challenge, limiting the broader application of TLS in large-scale studies of tree architecture.

In this study, we propose a novel algorithm, connectivity-based wood–leaf separation (CWLS), which integrates geometric classification with connectivity analysis to automatically extract wood points with high accuracy. To evaluate its performance and generalizability, we applied CWLS to TLS data collected from 55 trees representing diverse species and structural forms across six forest sites along a latitudinal gradient in eastern China, ranging from cold temperate to tropical zones. Each TLS point was manually annotated as ground truth.

CWLS achieved an average overall accuracy (OA) of 94.97%, precision of 93.34%, recall of 90.87% and F1-score of 91.97%. Notably, the algorithm maintained OA above 94.31% across all branch orders, demonstrating particularly strong performance in extracting wood points for higher order branches. Furthermore, CWLS outperformed three state-of-the-art wood–leaf separation algorithms—LeWoS, TLSeparation and graph-based leaf–wood separation—by offering a superior balance between precision and recall, especially for small branches in the upper canopy.

The ability of CWLS to substantially reduce noise while maintaining branch continuity makes it especially well suited for accurate and reliable 3D tree modelling from TLS data. Its integration with 3D reconstruction algorithms such as L1-tree and TreeQSM offers a promising pathway for large-scale quantification of tree architecture and for advancing our understanding of its adaptability and ecological functions under global climate change.