Individual tree segmentation from UAV LiDAR # based on PointNet++ network

Accurately identifying and extracting individual trees is an important foundational task in conducting forest inventory using remote sensing techniques. In this study, the common mixed coniferous and broadleaf forest in the northeastern region of China was selected as the study area, and the point cloud data was obtained using an unmanned aerial vehicle (UAV)-based light detection and ranging (LiDAR) system. In order to improve the accuracy of individual tree segmentation in mixed forest, an individual tree segmentation method based on the PointNet++ deep learning network was proposed. Firstly, a training dataset that included multi-channel features such as coordinates, RGB, intensity, and echo count was constructed. The channel attention mechanism and Inverted Residual Multilayer Perceptron module were added to the original PointNet++ network model to enhance the performance of the network. Then, the network was used to perform semantic segmentation on the coniferous trees, broadleaf trees, and ground points in the mixed forest. Based on the results of semantic segmentation, a method that combined local highest points and density maximum points to locate the individual trees was proposed and individual tree segmentation was finally achieved through a layer-stacking segmentation algorithm. Comparative analysis with three other traditional segmentation algorithms (Marker-based watershed algorithm, distance-based segmentation algorithm, mean shift algorithm, and PointNet) showed that the improved PointNet++ achieved the highest accuracy among the compared methods in this study ( r = 0.836, p = 0.849, F 1 = 0.843). Under the experimental conditions of this study, the proposed method improved the segmentation accuracy of individual trees in the mixed coniferous and broadleaf forest and can provide basic data for subsequent studies on tree structure parameter extraction, species identification, and biomass estimation.