TreeDBH: Dual Enhancement Strategies for Tree Point Cloud Completion in Medium–Low Density UAV Data

Medium–low density UAV point clouds often suffer from incomplete lower canopy structures and sparse distributions due to self-occlusion. While existing point cloud completion models achieve high metric accuracy, they inadequately address missing regions in trunks and lower canopy areas. To resolve these issues, this paper proposes a hierarchical random sampling strategy and a spatially constrained loss function. First, we dynamically stratify point clouds based on density distribution characteristics, employing hierarchical random sampling to preserve proportional representation of lower-level points, thereby effectively retaining basal tree structure information. Second, we introduce a distance constraint term for mid-lower point clouds into the symmetrical Chamfer distance (CD) loss, compelling models to prioritize completion quality in trunk base regions. Experiments on the FOR-instance-created completion dataset and Xiong’an dataset demonstrate that our method significantly enhances structural recovery capability at tree trunk bases, with visual results outperforming the baseline SeedFormer model. Additionally, we refer to existing point cloud-based diameter at breast height (DBH) calculation methods to measure the completed trees and compare the computed results with the measured values to evaluate the accuracy of the completion effect. Experimental results show that, after integrating our proposed strategies with existing completion methods, the accuracy of DBH measurement from point clouds is significantly improved. This study provides novel insights for addressing structural bias in tree point cloud completion and offers valuable references for digital forestry resource management.