LiDAR-Constrained ICP Post-Registration and UAV-SfM Accuracy Under Different Georeferencing Conditions in a Mixed-Surface Environment

Unmanned aerial vehicle Structure-from-Motion (UAV-SfM) products depend on acquisition and georeferencing. This study tested whether light detection and ranging (LiDAR)-constrained iterative closest point (ICP) post-registration reduces discrepancies in directly georeferenced products and improves digital surface model (DSM) agreement. Twelve configurations (three flight heights × four forward overlaps) were processed using ground control point (GCP)-constrained, directly georeferenced (noGCP), and LiDAR-ICP-registered (noGCP_ICP) strategies at a mixed-surface site; the noGCP models used onboard GNSS positions recorded as single-point (uncorrected) solutions. Reconstruction completeness, cloud-to-cloud (C2C) discrepancy, and the DSM of difference (DoD) were evaluated against the LiDAR data. Mean C2C root mean square error (RMSE) decreased from 2.934 m for noGCP models to 0.583 m after ICP, versus 0.353 m for GCP-constrained models. In the DoD, under the ±5 m sensitivity filter applied within each model’s own valid domain, mean RMSE in elevation (RMSEz) decreased from 2.988 to 1.300 m but remained above the 1.097 m of the GCP-constrained models. Improved point-cloud alignment therefore did not yield an equivalent reduction in DSM discrepancy. Because one LiDAR dataset supported both ICP estimation and evaluation, results represent LiDAR-relative rather than independent validation. LiDAR-constrained ICP is a conditional alignment strategy, not a substitute for robust georeferencing or image-network geometry. The reductions therefore quantify agreement with that reference and do not establish an independently validated improvement in absolute accuracy.