Automated detection and structuration of building and vegetation changes from LiDAR point clouds

Abstract. Urban environments are continuously changing, driven by factors such as population growth and infrastructure expansion, which necessitates regular updates to urban models. Accurate, up-to-date information on these changes is critical, particularly for national mapping agencies monitoring long-term urban development. This paper presents an automated methodology for detecting building and vegetation changes within urban environments using LiDAR point clouds, focusing on the city of Liège in Belgium. By leveraging recent aerial LiDAR data from 2022, our approach identifies, models, and integrates urban changes into a refined 3D Digital Twin model of Liège. The methodology includes preprocessing steps such as coordinate systems homogenization, noise filtering, and octree-based spatial indexing, followed by semantic and instance segmentation of point clouds using the RandLA-Net deep learning model. The change detection process focuses on four categories: appearance, disappearance, modification, and unchanged features. Achieving 100% accuracy for detecting new buildings changes, as validated within the study dataset and methodology. The modelled results are structured into a CityJSON city model. This automated approach significantly enhances urban model updates by integrating detected changes into a standardized 3D representation.