Monitoring Powdery Mildew With Hyperspectral Imagery In Wheat

Early monitoring of wheat powdery mildew is important for reducing the economic costs and environmental pollution caused by diseases.Accurate identification and early monitoring of diseases based on hyperspectral imaging are important research contents of crop growth monitoring.This study explored the application of near-ground imaging hyperspectral in the early monitoring of wheat powdery mildew.In this study,the single leaf and canopy field experiments were carried out based on different years,different varieties and different treatments,using imaging spectrometer to obtain hyperspectral imaging data of wheat healthy and infected leaves and healthy and infected canopy after inoculation with powdery mildew.At the leaf level and canopy level,subwindow permutation analysis(SPA)and continuous wavelet transform(CWT)were used to extract wheat powdery mildew sensitive spectral features and texture features.And construct normalized difference texture index(NDTI)and select spectral indices to compare with them.Then,build the wheat healthy and infected state recognition model and the disease estimation model based on Partial Least Squares-Linear Discrimination Analysis(PLS-LDA)and Partial Least-Squares Regression(PLSR),respectively,and to find out the accurate disease severity and time of the disease that could be distinguished at the earliest.At leaf level,the sensitive spectral features were extracted by SPA algorithm(533.5 nm,659.2 nm,528.5 nm,661.7 nm,968.2 nm,523.6 nm).By constructing NDTI it was found that the second moment(SEM)at the visible band of 659.2 nm had a good correlation with the disease severity(R2=0.64).The correlation between the other selected textures in NDTI and the disease severity were all lower than 0.55,but the correlation between the NDTI and the disease severity were significantly improved(R2≥0.59).The combination of spectral features and texture features is optimal and stable in PLS-LDA model and PLSR model.The overall classification accuracy of the discriminant model in all data dataset,jointing dataset and booting dataset were 82.29%,75.78%and 88.34%,respectively,and the Kappa coefficients were 0.62,0.47 and 0.75,respectively.The regression modeling R2 of all data dataset，jointing dataset and booting dataset were 0.82,0.65 and 0.89,respectively and the test R2 was 0.76,0.48 and 0.83,respectively.The daily discriminant model build by the combination of spectral features and texture features found that the jointing inoculation dataset was well recognized 6th day after inoculation with average leaf disease severity 3.9%and booting innoculation dataset was well recognized 4th day after inoculation with average leaf disease severity 6.2%,which was 2 days earlier than the leaf physiological parameter appears significant difference.Therefore,the combination of spectral information and texture information has a good application prospect in the early monitoring of crop diseases.At canopy level,the sensitive spectral features were extracted by CWT algorithm(595 nm,614 nm,708 nm,754 nm).By constructing NDTI,it was found that the mean(MEA)texture of 754 nm in the near infrared band had a superior performance in the texture index construction,and its correlation(R2)with the disease severity was 0.67.The correlation between the other selected textures in NDTI and the disease severity were all lower than 0.4,but the correlation between the NDTI and the disease severity were significantly improved(R2>0.4).Then it is found that the combination of wavelet features and texture features is the most accurate in the PLS-LDA model.The overall classification accuracy of the discriminant model is 81.17%,and the Kappa coefficient is 0.63.The PLSR model established by combining the spectral indices and the texture indices is the best,modeling and testing R2 were 0.76 and 0.71,respectively.Finally,based on the combination of wavelet features and texture features,the PLS-LDA model of different days after inoculation of wheat canopy powdery mildew was established.It was found that the earliest identification time of canopy wheat powdery mildew was 24-32 day after inoculation,and the earliest identified canopy average disease severity was 26-28%.The results show that texture information and spectral information can effectively improve disease identification and estimation accuracy.