一种激光雷达点云局部特征的树木枝叶分离方法  

A method of tree branch and leaf separation based on local features of LiDAR point

作　　者：张园园 孙颖[1,2] 张新长 ZHANG Yuanyuan;SUN Ying;ZHANG Xinchang(School of Geography and Planning,Sun Yat-sen University,Guangzhou 510006,China;Southern Marine Science and Engineering Guangdong Laboratory(Zhuhai),Zhuhai 519082,China;School of Geography and Remote Sensing,Engineering Technology Research Center of Intelligent Service for Urban and Rural Planning and Construction,Guangdong Province,Guangzhou University,Guangzhou 510006,China;College of Geography and Remote Sensing Science,Xinjiang University,Urumqi 830017,China)

机构地区：[1]中山大学地理科学与规划学院,广州510006 [2]南方海洋科学与工程广东实验室(珠海),珠海519082 [3]广州大学地理科学遥感学院,广东省城乡规划建设智能服务工程技术研究中心,广州510006 [4]新疆大学地理与遥感科学学院,乌鲁木齐830017

出　　处：《时空信息学报》2025年第1期73-82,共10页JOURNAL OF SPATIO-TEMPORAL INFORMATION

基　　金：国家自然科学基金面上项目(42171308);广东省自然科学基金面上项目(2024A1515012081);广州市基础与应用基础研究项目(202201011666)。

摘　　要：准确区分树木的枝干和叶片组分是获取叶面积指数等结构参数的关键;地面激光扫描仪(terrestrial laser scanning,TLS)可以获取毫米级点云,能够清晰刻画树木的精细结构信息,利用其点云进行分离枝叶以区分树木的枝干和叶片信息,是获取上述结构参数的基础。然而现有点云枝叶分离研究中常出现明显的误分现象,枝叶分离的精确度不高,因此本文提出一种基于点云局部特征的精细化点云枝叶分离方法。首先使用基于图的枝叶分离(graph-based leaf-wood separation,GBS)模型进行预处理,分离出初始的树木枝干和叶片点云;其次挖掘点云局部特征,利用曲率阈值、外接圆半径比较法分别对初始枝干和叶片点云进行精细化提取;最后为验证方法有效性,分别基于公开数据集和实验采集的数据,与随机森林算法、路径追踪检测算法和GBS模型进行比较。结果显示,本文方法的四种精度评价指标结果均表现为相对最优,其中,总体精度、Kappa值分别为0.945、0.811,分别比GBS模型的提高了0.027和0.072。[Objective]Accurately distinguishing between the wood(trunk)and leaf components of trees is crucial for obtaining structural parameters such as the leaf area index(LAI).Terrestrial laser scanning can capture point clouds at millimeter spatial resolution,allowing for detailed depiction of a tree's fine structure.The separation of wood and leaf point clouds from terrestrial laser scanning data serves as the foundation for deriving these structural parameters.However,current research methods exhibit significant misclassification in separating wood and leaf components.[Method]To enhance the accuracy of wood-leaf separation,this paper introduces a refined methodology leveraging local features within terrestrial laser scanning point clouds.Initially,a graph-based leaf-wood separation(GBS)model is employed for preprocessing to segregate the initial point cloud into branches and leaves.Subsequently,local features are extracted,and both branch and leaf point clouds are further refined using curvature thresholds and circumcircle radius comparisons.Given that local features pertain to neighborhood relationships,the study examines the effect of varying the number of neighboring points on separation performance,settling on 100 neighbors to balance computational efficiency and accuracy.Additionally,the choice of surface variation threshold is vital;here,the ratio of the maximum standard value(SV)to a variable parameterαis used as the segmentation criterion.Optimum results are achieved whenαis set to 1.45.For circumcircle radius comparisons,an empirical estimateγof 5 cm effectively purges misclassified wood points from the leaf point cloud.[Result]To validate the efficacy of our proposed method,we utilized both public datasets and self-collected data to acquire point clouds representing diverse tree structures.Comparisons against random forest,path tracking detection,and the GBS model reveal our approach yields the highest overall accuracy,Kappa coefficient,and F1 score,outperforming even the GBS model by 0.027 and 0.072 hig

关 键 词：地面激光扫描仪 点云 局部特征 曲率阈值法 外接圆半径比较法 枝叶分离 

分 类 号：P23[天文地球—摄影测量与遥感]