Deep Learning-Based Urban Fine-Grained Modeling and Geospatial Information Reconstruction Using Point Clouds

This paper presents a deep - learning - based method for urban fine - grained modeling and geospatial information reconstruction using point clouds. The proposed method first preprocesses point clouds through noise removal and data downsampling. Then, a two - stage hierarchical feature learning pipeline is employed to capture local details and global context. Urban components are segmented at semantic and instance levels, and geospatial information is reconstructed by 3D mesh generation and topological modeling. Experiments on public datasets (Paris - Lille - 3D and Semantic3D) and custom datasets (UAV LiDAR scans and terrestrial laser scans) show that the method achieves high performance in segmentation and reconstruction. For semantic segmentation, it reaches an mIoU of 84.5% on Paris - Lille - 3D and 76.1% on Semantic3D. In instance segmentation on custom UAV LiDAR data, the [email protected] for cars is 89.2%. The reconstruction results have a Chamfer Distance of 9.6mm, a completeness of 89.4%, and a normal consistency of 8.7°. Geospatial attribute accuracy is also satisfactory, with an RMSE of 0.85m for building height. The proposed method provides an effective solution for urban planning and analysis.