基于无人机多光谱影像的松材线虫病单木尺度监测

Tree Level Monitoring of Pine Wilt Disease Based on UAV Multispectral Imagery

松材线虫病是最具危害性的森林病害之一,亟需采取精准的监测手段来确定病疫木的株数和位置,实现松材线虫病的高效防控。利用多光谱无人机获取贵州省榕江县忠诚镇松材线虫病疫区图像,以无人机多光谱及其衍生点云作为数据源。首先,通过点云分割算法对研究区单木进行定位识别和树冠轮廓分割;然后,以分割单元提取光谱特征,并通过随机森林与递归特征消除相结合(RF-RFE)筛选出最佳特征集;最后,基于筛选特征集用于随机森林(RF)和支持向量机(SVM)检测模型构建,并评价模型检测性能,同时,使用RF和SVM对研究区进行感病情况反演,绘制松材线虫病空间分布图。结果表明:1)基于摄影测量点云单木分割效果较好,整体F-score为82.21%;经过特征筛选构建的RF模型,其OA和Kappa分别为84.4%和0.74,SVM的为76.09%和0.66。2)在树木健康、早期、中期和晚期4个阶段的检测中,RF的F-score值分别为78.43%,69.23%,83.33%和94.12%;SVM的为80.7%,55.81%,70.18%和84.13%。综合比较,RF的检测性能最好。研究表明,采用无人机多光谱影像和摄影测量点云相结合进行松材线虫病单木尺度监测具有可行性。通过研究,以期为低成本和精准的松材线虫病遥感监测提供参考。

Pine wilt disease(PWD)is one of the most harmful forest diseases,and there is an urgent need to adopt accurate monitoring means to determine the number and location of diseased trees for efficient prevention and control of PWD.In this study,the image of PWD epidemic area in Zhongcheng Town,Rongjiang County,Guizhou Province was obtained by using multi spectral UAV,and the multi spectral UAV and its derived point cloud were used as data sources.Firstly,localization identification and crown profile segmentation of individual trees in the study area were performed by point cloud segmentation algorithm.The spectral features were then extracted in segmentation units and the best feature set was filtered by a combination of random forest and recursive feature elimination(RF-RFE).Finally,random Forest(RF)and support vector machine(SVM)detection models were constructed based on screening feature sets,and the model detection performance was evaluated.At the same time,the RF and the SVM were used to invert the disease susceptibility in the study area and draw the spatial distribution map of PWD.The following key results were obtained:1)The individual trees segmentation based on photogrammetric point clouds was effective,with an overall F-score value of 82.21%.The OA and Kappa of the RF model constructed after feature screening were 84.4%and 0.74,respectively,and the SVM was 76.09%and 0.66.2)The F-score for RF were 78.43%,69.23%,83.33%and 94.12%,SVM were 80.7%,55.81%,70.18%and 84.13%for the four stages of tree health,early,middle,and late detection,respectively.The comprehensive comparison of the detection performance of RF was the best.The study pointed out that it was feasible to use the combination of UAV multispectral image and photographic measurement point cloud for individual trees scale monitoring of PWD.The study aimed to provide a reference for low-cost and accurate remote sensing monitoring of PWD.