Tree Species Classification of Point Clouds from Different Laser Sensors Using the PointNet++ Deep Learning Method

Abstract

Tree species information is a crucial factor in forest resource inventory. Light detection and ranging (LiDAR), as an emerging active remote sensing technology, has unique advantages in extracting three-dimensional (3-D) vegetation structure information, and its application in forest resource assessment and research is gaining increasing attention. Airborne laser scanning (ALS), unmanned aerial vehicle laser scanning (UAVLS) and terrestrial laser scanning (TLS) are important means to acquire 3-D forest data. The challenge of traditional machine learning based tree classification lies in extracting and selecting numerous key diagnostic features from large amounts of LiDAR data, requiring extensive feature extraction expertise, which limits its scalability. The use of deep learning methods for fast and accurate classification and identification of tree species in individual tree point clouds represents a new development direction of LiDAR technology in forest resource inventory applications. In this study, PointNet++ was used to classify tree species by point cloud data obtained from TLS, ALS and UAVLS, respectively. The research results show that high accuracy in tree species classification can be achieved by using point cloud deep learning methods.