Automatic LOD0 classification of airborne LiDAR data in urban and non-urban areas

Abstract

Point clouds are a very detailed and accurate vector data model of 3D geographic information. In contrast to other data models, no standard has been defined for visualization and management of point clouds based on levels of detail (LOD). This paper proposes the application of the concept of LODs to point clouds and defines the LOD0 for point cloud classification (the lowest possible level of detail) as urban and non-urban. A methodology based on the use of machine learning techniques is developed to perform LOD0 classification to airborne LiDAR data. Point clouds acquired with airborne laser scanner (ALS) are structured in grid maps and geometric features related with Z distribution and roughness are extracted from each cell. Six machine learning classifiers have been trained with datasets including urban (cities) and non-urban samples (farmlands and forests). The influence of grid size, point density, number of features and classifier type are analysed in detail. The classifiers have been tested in three case studies. The best results correspond to a grid size of 100 m and the use of 12 geometric features. The accuracy is around 90% in all tests and Cohen’s Kappa index reaches 81% in the best of cases.

Keywords

• Supervised classification
• machine learning
• point cloud
• big data
• laser scanner