Classify Tree Species from Point Clouds Generated by Different Laser Sensors: A Multi-View Projection Strategy

Accurately classifying forest tree species is crucial for monitoring forest resources and carbon storage. Light Detection and Ranging (LiDAR) is an emerging active remote sensing technology that provides higher spatial resolution and superior penetration capabilities for complex forest terrain and vegetation structures. Airborne laser scanning (ALS), uncrewed aerial vehicle (UAV)-borne laser scanning (ULS), and terrestrial laser scanning (TLS) are important methods for acquiring three-dimensional forest data. To address the issue of disorderliness inherent in directly using point cloud data for tree species classification, as well as the need to extract and select many key diagnostic features from a vast amount of lidar data for feature-based classification, this study employs a multi-view projection strategy and utilizes weighted average depth integration to incorporate depth information. The ResNet18/4 model is employed to extract and fuse features from depth images obtained from ALS, TLS, and UAVLS point cloud data for tree species classification. The research results demonstrate that the multi-view projection method achieves comparable or better results than the more complex state-of-the-art PointNet++ method, while being only half the size of PointNet++ and exhibiting better datasets generalization.