TLS-ULS point cloud registration in natural forests using hybrid semantic triangle matching and vertical structure weighting
Abstract
Establishing initial geometric correspondences between terrestrial laser scanning (TLS) and unmanned aerial vehicle laser scanning (ULS) forest point clouds is difficult under cross-platform observations. Fine registration is vulnerable to local non-rigid canopy deformation caused by canopy disturbance and platform-dependent observation differences. This study proposes a coarse-to-fine automatic registration framework that integrates hybrid semantic triangle matching and vertical structure weighting. Tree-top points and canopy-gap centroids are first extracted from the canopy height model (CHM). They are used to construct hybrid semantic triangle descriptors. The initial transformation is then estimated using random sample consensus (RANSAC) and singular value decomposition (SVD) for coarse registration. For fine registration, a normalized-height-based vertical structure weighting model increases the constraints from stable structures, including stems and the ground. It reduces the effects of canopy disturbance and cross-platform observational differences on iterative closest point optimization. Experiments were conducted in six Shangri-La plots with different slopes and canopy closures. The mean point-to-plane RMSE after coarse registration was 0.272 m, 27.8% and 23.2% lower than those of SDS (0.377 m) and TR (0.355 m), respectively. The mean point-to-plane RMSE after fine registration was 0.125 m. Relative to ICP, GICP, NDI, manual registration, and ForAlign, the proposed method reduced error by 47.8%, 23.5%, 47.5%, 26.9%, and 59.6%, respectively. These results show that the proposed method can align cross-platform point clouds in structurally complex natural forests. It provides a reliable basis for subsequent extraction of forest structural parameters.