A Two-stage Registration Method for UAV and HMLS Point Clouds in Subtropical Forest

Abstract

HMLS (Handheld Mobile Laser Scanning) and UAV (Unmanned Aerial Vehicle) LiDAR are increasingly utilized in forest inventory due to their efficiency and portability. However, challenges such as occlusions, low vertical overlap, and varying point cloud density complicate the fusion of these two datasets. In this study, we propose a novel two-stage method to match HMLS and UAV LiDAR data with different point density at complicate forest with dense canopy cover. The first stage optimizes voxel size selection for varying cloud densities and performs feature extraction. The second stage addresses gross error elimination through the truncated least squares method and performs feature matching using K-D Tree nearest neighbor indexing in combination with Singular Value Decomposition (SVD). The method was tested in 27 forest plots with varying vertical overlaps and stand conditions across Hubei Province, China and compared with four registration methods: Coarse-to-Global Adjustment Strategy (CGAS), Optimized Coarse-to-Fine Algorithms (OCFA), Generalized-ICP (GICP), and Bidirectional-Pearson Improved Method (BPIM). Results show that the proposed approach significantly improves registration accuracy, with error reductions of up to 0.096 m, 0.284 m, and 0.425 m under lower (0.37–0.56), moderate (0.58–0.73), and higher (0.77–0.95) canopy cover, respectively. Stand conditions and tree species influence registration accuracy. The results demonstrate higher accuracy in plots with lower canopy cover, steeper slopes, and fewer shrubs. Coniferous forests, with straighter trunks and fewer branches, provide more distinct feature points, leading to better accuracy than broadleaf forests. Additionally, UAV and HMLS matching accuracy is influenced by flight altitude, with higher altitudes increasing registration errors due to the decreased point density of UAV LiDAR.