Study On Remote Sensing Monitoring Of Pine Wilt Disease

Pine is one of the most important forest species in China,which has wide distribution,many species,good material and high economic value.Pine Wilt Disease(PWD),as known as "cancer of pine trees",has been spreading in China since it was found in 1982,which poses a great threat to pine resources,natural ecological environment and even national economy.Monitoring the epidemic area and situation of PWD is one of the key methods of control.Remote sensing technology has the characteristics of macroscopicity,timeliness,comprehensiveness and economy,which can provide a new way to monitor PWD.In this paper,non-imaging hyperspectral observation data and high spatial resolution remote sensing image of SuperView-1(SV-1)were used to monitor PWD at different scales.The main results are as follows:(1)Optimal Algorithm for Processing Hyperspectral Data of Pinus Massoniana.Continuous observation of the change of the spectral curve of pine canopy during the process from infection to death,and then based on four mathematical transformation methods:Savitzky-Golay filtering(SG),first-order derivative calculation(FD),inverse-log calculation(IL)and continuum removal(CR),band selection was carried out.Through the comparison of eliminate rate,T-test and Fisher discriminant accuracy,it was found that Successive Projections Algorithm(SPA)has high eliminate rate,fast speed and high discriminant accuracy in the form of band combination.(2)Band Window for Remote Sensing Monitoring of PWD.While Mean Confidence Interval(MCI)has higher significance and stability of discriminant accuracy in the single band,and the results are distributed continuously in interval form,so MCI results were used to band window analysis.The spectral data were divided into period ?,?,? and ? for analysis.In period I,only SG and IL form sensitive band window at 1300-1400nm and 1600-1900nm.Since period ?,460-505nm and 588-701nm can be effectively distinguished by SG,IL and CR.The sensitive band window of FD is relatively discrete,mainly distributed at 512-531nm.With the deepening of the degree of PWD,the sensitive band window is widened and the discrimination accuracy is improved.(3)Analysis of Optimal Model for Extracting PWD from Remote Sensing Images.SV-1 satellite,which has the highest spatial resolution,was selected to verify the band window in the middle and late period of the disease.Based on the spectral characteristics on UAV image,20 features,such as single band and vegetation index,were selected,which were evaluated according to the class separability,importance evaluation and redundancy between features.It was found that the normalized value and ratio of green band and red band of SV-1 remote sensing data are prior to other features.Combined with the training accuracy of the linear kernel,polynomial kernel,RBF kernel of SVM and RF classification model,the optimal classification model was selected.The results of the model show that SVM-RBF model is the best way to extract the epidemic area with RGVI,EVI,SR32,SR21 and NDVI,which recall rate is 87.72%,precision is 81.97%,and the F1-score is 0.8475.