Abstract
Airborne LiDAR point clouds represent terrain but are often sparse, noisy, and unevenly distributed, limiting their use in surface reconstruction and geospatial analysis. This study benchmarks four learning-based point cloud upsampling architectures, PU-Net, MPU, PU-GCN, and PU-Transformer, using real-world terrain data from the U.S. Geological Survey 3D Elevation Program (3DEP). A unified Farthest Point Sampling (FPS)-based preprocessing pipeline generates input and ground-truth patch pairs. Each model is trained with a fixed 4× upsampling ratio and input patch sizes of 128, 256, 512, and 1024 points. Performance is assessed on a held-out terrain sub-region using Chamfer Distance, Hausdorff Distance, Point-to-Surface Distance, Earth Mover’s Distance, and the Normalized Uniformity Coefficient. MPU achieves the best mean surface fidelity and point uniformity, obtaining the lowest Chamfer Distance, Point-to-Surface Distance, and Normalized Uniformity Coefficient at 1024 input points. PU-GCN achieves the lowest Hausdorff Distance and Earth Mover’s Distance at 512 input points. PU-Transformer remains competitive, whereas PU-Net produces consistently larger errors and does not benefit reliably from denser inputs. The three more recent architectures generally improve as patch capacity increases, indicating that richer local context improves terrain reconstruction. Overall, MPU and PU-GCN are the most suitable approaches for real-world LiDAR terrain point cloud upsampling tasks.