Deep Learning-Based Classification of Large-Scale Airborne LiDAR Point Cloud

摘要

Airborne LiDAR data allow the precise modeling of topography and are used in multiple contexts. To facilitate further analysis, the point cloud classification process allows the assignment of a class, object or feature, to each point. This research uses ConvPoint, a deep learning method, to perform airborne point cloud classification at scale, in rural and urban contexts. Specifically, our experiments are located near Montreal (QC) and Saint-Jean (NB) and our approach is designed to classify five classes; we used “Building”, “Ground”, “Water”, “Low Vegetation” and “Mid-High Vegetation”. Experimenting with different configurations, we achieved excellent Intersection-over-Union results for the “Mid-High Vegetation” (93%) and “Building” (86%) classes on both datasets and provide insights to improve processing times as well as accuracy.