A Tree Species Classification Model Based on True-Color Point Clouds From a UAV-Based Multisensor System

Abstract

Accurate tree species identification is central to sustainable forest management, biodiversity conservation, and ecological monitoring. However, due to the high density and nonuniform spatial distribution characteristics of point clouds acquired by uncrewed aerial vehicle-based light detection and ranging (UAV-LiDAR), existing deep learning models generally encounter challenges such as high inference latency and the loss of local geometric signals when processing such sensing data. To address these issues, this article proposes an efficient sensing architecture named DL-PointNet++, designed for high-precision and high-efficiency tree species classification of multimodal sensing data in complex forest environments. Built upon the PointNet++ framework and tailored to the sampling characteristics of LiDAR sensors, the architecture introduces a density-aware voxel point sampling (DAVPS) strategy. This approach mitigates the nonuniform spatial distribution of raw point clouds at the algorithmic level, significantly enhancing sampling efficiency while maintaining high-fidelity structural representation. To compensate for signal attenuation during deep feature extraction, a local feature enhancement (LFE) module is incorporated to strengthen the model’s ability to capture fine-grained crown details. Furthermore, the focal loss function is employed to optimize loss weights, addressing the inherent class imbalance in field-measured data and improving the system’s recognition robustness across diverse scenarios. Additionally, a new high-resolution dataset for tree species classification was constructed and utilized for validation. Experimental results demonstrate that DL-PointNet++ achieves superior classification performance while maintaining an acceptable computational cost, providing a practical framework for efficient tree species classification based on UAV-derived true-color point clouds.