Remote Sensing-Based Identification Of Wheat Powdery Mildew Based On Sensitive Spectral Characteristics

Powdery mildew is one of the most important diseases that endanger the yield and quality of wheat.However,the previous monitoring methods were mainly based on visual inspection,qualitative description and subjective judgment.The observation methods with point-to-face often delay the timely prevention and control of the disease and increase the difficulty and cost of post-governance.The development of remote sensing technology had made it possible to monitor diseases at multiple scales,such as leaves,canopies and regions.The purpose of this paper is to study the ability of sensitive traits of wheat powdery mildew for monitoring the severity of powdery at different scales in two situations including noise or not.Based on this goal,non-imaging hyperspectral,satellite multispectral,ground measured,digital elevation model(DEM),thermal infrared and other data of different scales and sources were synthesized,and the sensitive features of wheat powdery mildew severity based on leaf hyperspectral data and simulated canopy data were selected,respectively.At the same time,the land surface temperature(LST)calculated by Landsat-8 and MODIS was used as the key habitat feature and supplemented by the selected vegetation index as the growth factor,thus,the effect and feasibility of grading the severity of wheat powdery mildew was studied by the performance of the combination data on the stars and the ground.The important research contents of this paper are as follows:(1)Study on the differentiation method of powdery mildew severity based on non-imaging hyperspectral technique.A total of 19 hyperspectral features based on hyperspectral position characteristics and common vegetation indices characteristics were selected.The relationship of wheat powdery mildew with leaf hyperspectral sensitive features were studied and the spectral curves of wheat leaves in the early stage of grain filling were selected as data source.In the screening of variables,the significance of the correlation between hyperspectral features and wheat powdery mildew severity was verified by independent sample T test;then,the linear regression was performed based on the selected features.The regression models established by most of the hyperspectral features performed well.In order to further improve the ability of regression model for monitoring the severity of powdery mildew,according to the variable cross-correlation matrix,four features(MSRg,NDVIn,?g,SIPIn)were selected for multivariable linear regression,respectively.The results showed that the combination of multiple independent variables could effectively improve the linear regression model to distinguish powdery mildew severity;in addition,disease grades of powdery mildew were classified based on random forest algorithm with classification accuracy of 82.7%and Kappa coefficient of 0.73.(2)Stability assessment of sensitive features to disease monitoring in simulated noise environments.Based on the non-imaging hyperspectral data of winter wheat leaves,the robustness of sensitive features on simulated background noises was validated.The four sensitive hyperspectral features(MSRg,NDVIn,?g,SIPIn),which were verified by variable screening and regression analysis,were selected.By changing the type and weight of input noise,the correlation between each feature and the severity of powdery mildew was studied.The results indicated that the addition of noises would affect the degree of correlation between each feature and powdery mildew severity.Subsequently,under combined noise interference(soil,clouds and trees)conditions,the monitoring ability of GF-1 WFV and Landsat-8 OLI data simulated by their respective band ranges and spectral response functions for wheat powdery mildew was further explored.Seven indicators including basic bands(blue,green,red and near-infrared),normalized vegetation index(NDVI),normalized greenness vegetation index(NDGI)and pigment-sensitive vegetation index(SIPI)were analyzed,respectively.The results proved that GF-1 and Landsat-8 had strong monitoring ability for powdery mildew with interference from background noise;GF-1 could be used as a reliable data source for wheat powdery mildew monitoring.(3)Remote sensing monitoring of powdery mildew on regional scale combined with vegetation indices and land surface temperature(LST).Jinzhou City,Hebei Province,often diagnosed with severe wheat powdery mildew in northern China,was selected as a research area to study the remote sensing monitoring for powdery mildew on regional scale.Vegetation indices calculated by GF-1 WFV data which represented wheat growth and LST(calculated by Landsat-8 and MODIS data)established a classification model based on support vector machine algorithm(suitable for small sample classification),and then,wheat powdery mildew on regional scale was monitored by this model.We used Relief-F algorithm to obtain sensitive variables in vegetation index screening.There were four types of LST:single-time Landsat-8 LST,multi-temporal Landsat-8 LST,multi-temporal MODIS LST and LST which calculated by Lansat-8 and MODIS data.These four LST combined with the same vegetation indices were used to calculate the severity of wheat powdery mildew.The results indicated that the last model had the highest accuracy,followed by the second one,the first one.The accuracy of these four models is 81.2%,76.8%,73.9%and 64.8%,respectively.This indicated that land surface temperature is one of the most key factors affecting the severity of wheat powdery mildew.The addition of multi-temporal surface temperature data significantly improved the accuracy of wheat powdery mildew monitoring.In summary,the remote sensing monitoring of wheat powdery mildew could be put into practical application.The next step is to explore the ability of land surface temperature and other factors for monitoring the severity of powdery mildew timely,accurately and extensively.