Enhanced 3D reconstruction of tall buildings from UAV photogrammetry using GCP clustering and multi-scale evaluation

High-quality three-dimensional (3D) point clouds are essential for representing the geometry of structures in smart-city applications. To obtain accurate point clouds of tall buildings, this study develops a clustering-guided framework for designing ground control point (GCP) configurations and performing multi-scale accuracy evaluation in unmanned aerial vehicle (UAV) photogrammetry. This framework adopts K-means clustering to group candidate GCPs and utilizes the resulting clusters to guide reconstruction experiments with various numbers and spatial distributions of GCPs. Reconstruction accuracy is assessed at the point, line, and surface levels. This study validates the proposed method using a high-rise building and a super-tall tower in China. Incorporating clustering algorithms reduces reconstructed errors by up to 22.2%, as compared to those without clustering. The results also reveal that a sufficiently large number of GCPs is required to obtain satisfactory 3D models, and a uniform distribution of GCPs yields the lowest reconstructed errors. Centimeter-level precision can be achieved in point coordinates and line lengths, with root-mean-square errors (RMSEs) of less than 33 mm and 40 mm for the high-rise building, and those of less than 40 mm and 61 mm for the super-tall tower, respectively. The proposed method reduces the RMSEs in surface fitting by more than 7% and 20% for the high-rise building and super-tall tower, respectively, compared to cases without GCP. Overall, combining clustering-guided GCP configuration design with multi-scale accuracy evaluation provides a quantitative basis for analyzing the effects of GCP number and spatial distribution on the reconstruction accuracy of the examined tall buildings.