From Wood–Foliar Semantics to 3D Fuel Characterization: Transferable Deep Learning for Terrestrial LiDAR Across Global Forest Ecosystems

Terrestrial Light Detection and Ranging (LiDAR) provides detailed three-dimensional (3D) observations of forest structure, yet transferable wood–foliar semantic segmentation remains challenging because of forest structural heterogeneity, occlusion, and variability across forest ecosystems and terrestrial LiDAR platforms. Existing Deep Learning (DL) approaches are commonly developed for localized forest conditions or complex multi-class semantic taxonomies, often limiting their transferability to structurally distinct forests. Here, we introduce points2SBL, a geometry-driven, terrestrial LiDAR platform-agnostic DL framework that reduces heterogeneous forest vegetation to a transferable wood–foliar representation and subsequently uses these semantics as the foundation for 3D forest fuel characterization. The framework was developed using globally harmonized benchmark datasets acquired across multiple terrestrial LiDAR systems and broadleaf, coniferous, mixed, regenerating, and structurally complex forests. Three point-based architectures (PointNet++, PointNeXt, and Point Transformer) were benchmarked under identical training and evaluation protocols. Point Transformer provided the most consistent cross-dataset performance, achieving overall accuracy (OA) of 92–94%, mean Intersection-over-Union (mIoU) of 0.78–0.87, macro F1-score of 0.87–0.93, and Matthews correlation coefficient (MCC) of 0.74–0.86 across five independent benchmark datasets. Species-level, vertical-profile, and qualitative cross-ecosystem assessments further demonstrated consistent preservation of wood–foliar semantics throughout the 3D forest fuel continuum, including previously unseen forest point clouds, while revealing that some apparent semantic disagreements originated from manual annotation of benchmarks. Finally, the points2SBL wood–foliar semantics were geometrically decomposed using a geometric algorithm named “Points2woodyseg” into fuel-related components: stem, branches, leaf, surface wood, and surface foliage. Collectively, these results demonstrate that a simplified wood–foliar abstraction can provide a transferable semantic foundation for characterizing structurally heterogeneous terrestrial LiDAR point clouds and deriving essential 3D forest fuel information across diverse forest ecosystems.