Single Tree Crown Segmentation Based on Rasterization of UAV LiDAR Point Cloud and Mask R-CNN

This study aimed to employ deep learning algorithms for accurate single-tree crown segmentation from point cloud data, addressing the critical issue of single-tree segmentation in the process of forest structure parameter extraction. The LiDAR point cloud from unmanned aerial vehicle (UAV) was a common data source for single tree segmentation; however, the point cloud data was large, and the processing was complex. To reduce the difficulty of point cloud data processing and improve the accuracy of single tree segmentation in complex stands, point cloud rasterization processing combined with deep learning algorithm was proposed to segment the canopy of single tree in this study. Firstly, the Feima Robotics company’s D2000 UAV platform equipped with the D-LiDAR2000 LiDAR sensor was used to acquire the point cloud data of a mixed coniferous and broadleaf forest. Lidar360 software was employed for point cloud denoising, ground point classification, and point cloud normalization preprocessing. Subsequently, the top-down rasterization of the sample plot cloud was performed, and the maximum height, maximum intensity , and density information per unit rasterized area of the point cloud were calculated, which were then mapped to the RGB channels corresponding to the rasterized image pixels, making the tree canopy clearer in the rasterized image. Secondly, according to the Mask R-CNN model within the Detectron2 framework, the number of layers and iteration of different backbone networks were compared to select a backbone network that provided superior segmentation performance. Then, the Global Context Network (GC) and Attention Mechanism modules were integrated into the ResNet network, and compared with the simultaneous introduction of the GC Net and Attention Mechanism modules to enhance the segmentation accuracy of the Mask R-CNN model. To validate the practicality of the improved Mask R-CNN model, its segmentation accuracy was compared with that of similar deep learning networks(U-Net and DeepLabv3+). Finally, the tree crown masks segmented by the improved Mask R-CNN were used to segment the point cloud of individual tree crowns. The segmentation results of the test plot were compared and evaluated using the watershed algorithm, K-means algorithm, and the improved Mask R-CNN algorithm, respectively. Among the three backbone networks of Mask R-CNN, the R50-FPN-3X network saved some training time and computational resources compared to the R101-FPN-3X network, with an average accuracy of 76.72%, increasing by 1.01 percentage points higher than that of the R50-FPN-1X network. Simply introducing the Squeeze-and-Excitation (SE), Coordinate Attention (CA), and Convolutional Block Attention Module (CBAM) mechanisms, the model accuracy increased to 78.13%, 78.86%, and 81.37%, respectively, increasing by 1.41, 2.14, and 4.65 percentage points, respectively. After integrating the GC Net module into the ResNet network, the model accuracy was 80.35%, which is an improvement of 3.63 percentage points over the original model. While introducing both of the CBAM and GC Net attention mechanisms achieved the highest accuracy of 82.91%, increasing by 1.54 percentage points and 2.56 percentage points compared to using the two modules alone. The improved Mask R-CNN model achieved an average accuracy that was 7.27 percentage points higher than the U-Net model and 4.62 percentage points higher than the DeepLabv3+ model. The improved Mask R-CNN algorithm outperformed the Watershed and K-means algorithms in single-tree crown point cloud segmentation, achieving the highest recall, precision, and F-score of 81.19%, 78.85%, and 80%, respectively. The single-tree crown point cloud segmentation method based on the improved Mask R-CNN network demonstrated robustness in tree crown segmentation in mixed coniferous -and broadleaf forests. By integrating point cloud data processing with deep learning models, it enhanced the accuracy of single-tree crown segmentation, thereby providing reliable foundational data and technical references for forest resource assessment, biomass estimation, and carbon stock estimation.