Monitoring Powdery Mildew With Hyperspectral Reflectance In Wheat

The powdery mildew has risen to one of the main wheat production diseases and presents an increasing trend of spreading year by year,which has caused serious losses to wheat production in Jiangsu province and even in whole country.Therefore,it is urgent to develop a rapid and accurate technology to identificate and monitor powdery mildew.Hyperspectral remote sensing is one of the non-destructive technologies that provides a possible method to solve this problem.In this study,pot experiment was conducted to collect agronomic parameters of wheat leaves such as hyperspectral reflectance,chlorophyll content and net photosynthetic rate based on natural infection conditions by powdery mildew at late jointing period with two different resistant wheat varieties ’Yangfumai 4’ and ’Shengxuan 6’.The hyperspectral features indicating the growth status at preinfection were extracted by Subwindow Permutation Analysis(SPA)in order to establish the model based on Partial Least Squares-Linear Discrimination Analysis(PLS-LDA).The hyperspectral features indicating the leaf disease severity were extracted by methods of Competitive Adaptive Reweighted Sampling(CARS),Uninformative Variable Elimination(MC-UVE),Random Frog(RF)respectively,and the models were established by Partial Least Squares Regression(PLSR).The expected results not only provide core band selection for the development of crop diagnosis instrument,but also provide effective technical support for monitoring crop disease severity with hyperspectral remote sensing,so as to promote the development of disease research in information agriculture.Firstly,leaves of two different resistant wheat varieties were studied to analyze the dynamic changes of physiological and ecological parameters and the spectral reflectance of wheat leaves at the early and middle stage of powdery mildew.The sensitive range of wavebands of wheat disease were identified and confirmed in order to further comparing the sensitivity of the existing spectral indices to disease severity between the two varieties.The experiment results show that pigment content and photosynthetic capacity of the two varieties were all decreased rapidly with the disease severity increasing while changed slightly at the early stage.Red and near-infrared region gave first response to the disease at the early stage and the sensitive bands range of ’Shengxuan 6’ was wider than the one of ’Yangfumai 4’.The visible spectral region all gave response to the disease at the middle stage.Therefore,it can provide a substantial theoretical basis for disease diagnosis when the dynamic changes of physiological and ecological parameters and hyperspectral reflectance are studied explicitly.Secondly,sensitive bands,hyperspectral features extracted from reflectance by SPA,hyperspectral features extracted from spectral indices by SPA were employed as input variables of PLS-LD A respectively to establish and compare the models of identifying wheat health status.The results show that the accuracy of PLS-LDA model constructed with 316 sensitive bands was 98.21%,but the test accuracy of ’Shengxuan 6’ was only 65.27%.The accuracy of PLS-LDA model constructed with 12 features extracted by SPA from spectral reflectance was 85.12%,the cross validation accuracy was 84.52%and the test accuracy of’Shengxuan 6’ was 84.43%,and the accuracy of PLS-LDA model constructed with 4 features extracted by SPA from spectral indices was 82.14%,80.95%and 85.63%respectively.The research results also point out that the sensitive bands are only suitable for detecting health status of a single variety while SPA can produce less variables with higher stability and reliability of the model.In conclusion,it’s simple to detecting disease when using spectral reflectance directly.Finaly,vegetation index method and PLSR method were used to monitor disease severity of wheat leaves.The quantitative relationship between existing spectral indices and Normalized Difference Spectral Index(NDSI)to disease severity were studied and compared.The hyperspectral features were extracted using CARS,MC-UVE and RF algorithm respectively,and the corresponding PLSR models were compared.The performance of model accuracy and error control established by NDSI(R492,R451)and NDSI(R643,R433)were better than existing spectral indices.Among this four kinds of methods,RF and CARS could produce less variables and less running time than MC-UVE,moreover,the performance of models based on RF and CARS were better than MC-UVE and vegetation index method which contain less spectral information.In conclusion,this study will provide guidance for helping monitor wheat diseases through method of remote sensing in future.