Unmanned Aerial Vehicle Light Detection and Ranging (UAV-LiDAR) has emerged as a powerful tool for capturing three-dimensional forest information, enabling detailed analysis of tree structures and compositions. Yet application of individual tree segmentation (ITS) from UAV-LiDAR in Japanese forests remain challenging. Using UAV-LiDAR data acquired over both a complex natural forest and a cedar plantation in northern Japan, we evaluated a recent deep learning-based ITS approach (ForestFormer3D) for individual tree detection and structural attributes estimation.
Results show that segmentation performance varies with canopy complexity and stand composition. In the natural forest plot (100 × 50 m), the method detected up to 46% (119 out of 257) of field-recorded trees, largely because the stand was dominated by small trees (~69% with diameter < 20 cm). Of the detected segments, 14 represented multi-stem clusters of Birch and other broadleaf species, which are difficult to segment individually. Successfully detected individuals were predominantly medium and large size classes (diameter >20 cm, height >10 m). Estimated crown areas of segmented trees showed good agreement with field measurements (r = 0.70), while tree height estimates were stronger (r = 0.97). Segmented crown area demonstrated a moderate correlation with the stem diameters (r = 0.60). In dense plantation plot (80 × 80 m), segmentation accuracy was substantially higher, with 495 trees detected and only 27 individuals not clearly separable. Although field validation was unavailable for this plot, manual inspection confirmed generally reliable tree separation with a homogeneous stand structure.
Overall, UAV-LiDAR combined with ITS approach enables reliable reconstruction of individual tree structure, particularly for medium and large trees that account for a disproportionate share of forest biomass and productivity. Because the method directly derives spatially explicit tree-level attributes (height and crown size), it provides structurally meaningful inputs for ecosystem assessment. The UAV-based workflow further enables efficient extension from field plots to landscape-scale coverage, making this approach practically suitable for large-area ecosystem monitoring.