应用堆叠融合模型和点云数据的树种分类

Tree Species Classification Using Stacking Fusion Model and Point Cloud Data

2021年6月份,以福州市三江口生态公园为研究区域,应用科卫泰的六旋翼无人机KWT-X6L-15搭载RIEGL VUX-1UAV三维激光扫描仪采集数据,获取8种共计444棵优势树种点云数据;使用LiDAR360激光雷达点云数据处理分析软件对激光雷达(LiDAR)点云数据预处理后,进行单木分割操作,提取树木参数,筛选得到树高、地径、枝下高、冠幅面积、冠高、胸径、95%百分位高度7个单木特征参数;使用堆叠(Stacking)融合模型应用点云数据提取的参数因子进行树种分类,分类结果与常用的支持向量机、随机森林、K最近邻模型3种模型分类结果进行对比,分析堆叠融合模型应用点云数据进行树种分类的准确度、卡帕(Kappa)系数、精确率、召回率等。结果表明:堆叠融合模型的树种分类,准确度达79.72%、卡帕系数为0.7681;支持向量机、随机森林、K最近邻模型3种模型的分类,准确度分别为60.14%、67.57%、63.95%,卡帕系数分别为0.5443、0.6327、0.5829;堆叠融合模型对树种分类的效果,整体优于K最近邻模型、随机森林模型、支持向量机模型3种常用模型。堆叠融合模型对水杉(Metasequoia glyptostroboides)、小叶榕(Ficus concinna)、朴树(Celtis sinensis)、南洋杉(Araucaria cunninghamii)的识别效果更好,对水杉分类精确率达到90.91%,对小叶榕、朴树、南洋杉的分类精确率也能达到80%。不同树种分类模型对同一种类树木的分类精度存在明显差异,在所选试验树种中堆叠融合模型适用性更高。

In June 2021,Sanjiangkou Ecological Park in Fuzhou City was selected as the research area,and the data were collected by KWT-X6L-15,a six-rotor UAV equipped with RIEGL VUX-1UAV 3D laser scanner,and the point cloud data were obtained for a total of 444 dominant tree species of eight species.After pre-processing the LiDAR point cloud data using LiDAR360 LiDAR point cloud data processing and analysis software,single wood segmentation operation was performed to extract tree parameters.The tree height,ground diameter,branch height,crown area,crown height,diameter at breast height and 95%percentile height were obtained.The accuracy,Kappa coefficient,precision and recall,etc.of the stacked fusion model for tree classification using point cloud data were analyzed.The results showed that the accuracy of the stacked fusion model for tree species classification was 79.72%and the kappa coefficient was 0.7681;the accuracy of the three models of support vector machine,random forest and K-nearest neighbor model were 60.14%,67.57%and 63.95%,respectively;the kappa coefficients were 0.5443,0.6327 and 0.5829,respectively;and the overall effect of the stacked fusion model on tree species classification was better than the three commonly used models of K-nearest neighbor model,random forest model,and support vector machine model.The stacked fusion model was better than the K-nearest neighbor model,random forest model,and support vector machine model for the classification of Metasequoia glyptostroboides.