TreeSemantic: An Online Module for Individual Tree Segmentation and Attributes Estimation

Abstract

The development of laser scanning (LS) systems and simultaneous localization and mapping (SLAM) algorithms has rapidly advanced in recent years. Their integration enables the efficient capture of the 3-D structure of forests, but the attribute estimation at the individual tree level typically requires postprocessing in professional software. To address the limitation, this study presents TreeSemantic, an online individual tree segmentation and attributes estimation algorithm. TreeSemantic is designed as an online framework tightly coupled with light detection and ranging (LiDAR)-SLAM. It introduces an equivalent projection strategy that converts irregular 3-D point clouds into structured 2-D representations, allowing efficient normal estimation and individual tree segmentation. In addition, a spatiotemporal aggregation mechanism with external and internal consistency verification is developed to incrementally reconstruct individual trees from sequential scans, mitigating incomplete stem observations caused by occlusion and limited viewpoints. Based on the continuously refined tree clusters, the tree location, diameter at breast height (DBH), and stem curve are estimated in real time. Moreover, commonly used forest inventory products, such as the digital elevation model (DEM), tree attributes database, and pointwise labeled data (tree and ground), are generated immediately after TreeSemantic ends. Extensive experiments were conducted on the self-collected Wuhan University (WHU) dataset and the public TreeScope dataset. The experimental results demonstrate that TreeSemantic achieves nearly 100% tree detection and centimeter-level accuracy in tree attribute estimation. More specifically, the DBH RMSE ranged from 3.82 to 7.93 cm, with the relative RMSE from 15.01% to 29.35%. The stem curve evaluation on 137 matched trees yields an average RMSE of 4.82 cm and a Bias of 1.13 cm. Furthermore, the average running times of normal estimation, segmentation, spatial–temporal aggregation, and attribute estimation are 8.52, 23.46, 13.06, and 11.32 ms, respectively, demonstrating millisecond-level efficiency. Comparisons with state-of-the-art methods, including SLAOM, RealtimeTrees, and 3DFin, further demonstrate that TreeSemantic achieves a favorable balance between accuracy, robustness, and real-time capability. Overall, TreeSemantic provides a practical and efficient solution for users, such as farmers, forest managers, or people without any professional knowledge, to obtain individual tree attributes directly during data acquisition without additional postprocessing.