3D Convolutional Autoencoder For Tree Extraction From Mobile And Airborne Lidar Point Clouds

Zaboli, Mahdiye; Shokri, Danesh; Dolati, Fariba; Homayouni, Saeid

doi:10.22059/eoge.2023.357160.1136

3D Convolutional Autoencoder For Tree Extraction From Mobile And Airborne Lidar Point Clouds

Document Type : Original Article

Authors

¹ Department of Photogrammetry and Remote Sensing, School of Surveying and Geospatial Engineering, College of Engineering, University of Tehran, Tehran, Iran

² Department of Photogrammetry and Remote Sensing, School of Surveying and Geospatial Eng., University of Tehran

³ MSc photogrammetry at Tehran university

⁴ Centre Eau Terre Environnement, Institut national de la recherche scientifique (INRS), Quebec G1K 9A9, Canada

10.22059/eoge.2023.357160.1136

Abstract

Tree information is essential in urban planning and green space resource management to mitigate air pollution, produce oxygen, and preserve wildlife and biodiversity. Current tree inventory in urban areas is done by in-person field data collection and takes several years to complete. However, this data are relatively accurate and high-resolution; it covers limited spatial coverage. Recent progress in remote sensing technologies and methodologies provides promising opportunities for accurate and scalable tree mapping and monitoring. In particular, deep learning (DL) models to analyze the high potential 3D observations from airborne, terrestrial, and mobile LiDAR systems helped in reliable and cost-effective urban tree inventory. Although LiDAR point clouds comprise valuable 3D information, most proposed methods convert 3D input data to 2D projections for exploiting DL methods. This paper presents a new tree extraction method based on a 3D convolutional autoencoder (3D-CAE) applied directly to LiDAR point clouds. Initially, ground point filtering was used to remove unnecessary points. Then, the designed model was trained to encode deep features automatically. Finally, the Support Vector Machine (SVM) was employed to classify tree-related point clouds. Experiments on five datasets, including mobile laser scanner (MLS) and airborne laser scanner (ALS) data, illustrated the efficiency of the proposed method. Overall, a 95% Precision demonstrated the suggested method’s high performance in individual tree detection and identification. In addition, the algorithm computed the geometrical information of classified trees, such as 2D coordinate space, height, stem diameter, and 3D boundary tree locations.

Keywords