Early Detection Of Fusarium Head Blight In Wheat With Ground-based Hyperspectral Data And Machine Learning

Wheat fusarium head blight(FHB)is a devastating fungal disease that poses a serious threat to crop yield and quality.Therefore,early detection and accurate diagnosis are important for the selection of resistant varieties,accurate spraying,and integrated disease control measurements.Traditional crop disease monitoring and detection requires a large amount of human and intensive resources with potential issues such as higher error and low efficiency.However,emerging imaging and non-imaging hyperspectral technologies and machine learning methods offer the possibility of real-time monitoring and accurate diagnosis of crop diseases.In this study,we obtained temporal imaging and non-imaging hyperspectral data,disease severity and biochemical parameters of healthy and FHB infected wheat at two scales: spike and canopy.Two winter wheat varieties inoculated with FHB for three consecutive years were examined for the spectral variation patterns of the spike and canopy scales under different disease severity scales.Afterwards,the sensitive features to FHB were selected based on continuous wavelet transform(CWT)and random forest-recursive feature elimination(RF-RFE)methods and examined using five machine learning methods(K-nearest neighbor(Knn),Random forest(RF),Support vector machine(SVM),Neural net(NN),Extreme gradient boost(Xgboost))for healthy and susceptible wheat classification and disease severity estimation.The results of the study provide operational and systematic references for early and midterm monitoring of wheat FHB disease and provide technical support for accurate application.Firstly,hyperspectral imaging of wheat ears was used as a data source to explore the potential of early monitoring of FHB.Four sensitive wavelet features(WFs-WF423,WF581,WF624 and WF865)were extracted based on the CWT,and the texture features of the corresponding wavelet coefficients were calculated for each of the eight grey-scale co-occurrence matrices(Mean,Variance,Homogeneity,Contrast,Dissimilarity,Entropy,Second Moment,and Correlation),a total of 32 texture features were input into an RFRFE model.The five texture features(COR423,MEA581,HOM865,CON865 and DIS865),that were sensitive to the disease in all years were filtered out.The new wheat fusarium spectral indices(WFSI)and textural indices(WFTI)were developed and evaluated by machine learning methods(Knn,RF,SVM,NN,Xgboost)for disease classification.The newly developed spectral indices(WFSI1 and WFSI2),and textural indices(WFTI1 and WFTI2),individually manifested CA up to 79% for pre-symptomatic.The developed indices WFSI1,WFSI2,WFTI1 and WFTI2,individually manifested average classification accuracy(ACA)in all machine learning classifiers(MLCs)of 78.90,72.0,73.60 and71.0%,respectively at the pre-symptomatic scale which increased to 90.10,90.30,88.10 and 85.40% ACA,respectively at DS2(3 to 5% disease percentage(DP)).Furthermore,the fusion of all four developed indices in the years 2019,2020 and 2021 showed improved ACA of 79.05,76.75 and 78.59%,respectively for pre-symptomatic that increased to 98.06,89.68 and 94.83% ACA,respectively at DS2.Among the five MLCs,Xgboost-based severity prediction performed the best(accuracy R2 = 0.92,error RMSE =7.87),followed by the results of Knn.This part of the study reveals the promising implementation of hyperspectral spatial and spectral information for improving the FHB early monitoring in precision agriculture applications.Secondly,the hyperspectral reflectance of wheat ears was used as a data source to explore the potential of non-imaging spectroscopy for early monitoring and diagnosis of FHB.Disease-sensitive features were selected from 34 conventional sensitive indices and2 newly constructed spectral indices(WFSI1,WFSI2)according to a variable importance score(VIP)method.Subsequently,the machine learning-sequential floating forward selection(ML-SFFS)method was used to construct healthy and disease-sensitive classification models.Finally,the combination of disease-sensitive features was used to construct a quantitative inverse model of wheat FHB disease severity.The results showed that a multivariate subset(NDVI,RR,WFSI1 and WFSI2)coupled with the newly constructed WFSI1 and WFSI2 were the most sensitive features for monitoring wheat FHB,with an average classification accuracy(ACA)of 82.67% at pre-symptomatic(DS1).At the DS2 stage,the mean ACA reached 99.3%.Seven previously published indices such as SRPI,PSNDa,NDVI,RR,PSSRb,RARSb and CAR also performed better in identifying healthy and susceptible plants.In addition,the results of the quantitative inversion model of wheat FHB disease severity based on a multivariate subset of Knn outperformed other disease models built by machine learning.The results of this study suggest that machine learning methods such as SVM and Knn based on nonimaging reflectance combined with ML-SFFS,have good potential for plant health and disease susceptibility classification and estimation.Finally,the potential of non-imaging spectroscopy for early monitoring of wheat FHB was explored using wheat canopy hyperspectral reflectance as a data source.Based on the elucidation of spectral response patterns following disease invasion,five MLCs(Knn,SVM,RF,NN and Xgboost)were applied to classify healthy and susceptible plants for wheat FHB based on the previous construction of disease-sensitive vegetation indices and two newly constructed spectral indices(WFSI1 and WFSI2).In addition,compared univariate and multivariate linear regression methods to construct canopy reflectance performance for disease severity estimation models.The results showed that the spectral reflectance of wheat canopies inoculated with the disease increased in the visible region,decreased significantly in the near-infrared region,and increased significantly in the short-wave infrared region.Five sensitive spectral bands(401 nm,460 nm,570 nm,786 nm,and 840 nm)were extracted using CWT and a canopy level wheat fusarium canopy indices(WFCI1 and WFCI2)was constructed and screened based on the above sensitive spectral bands.The newly constructed canopy level disease indices WFCI1 outperformed the other indices for monitoring and diagnosing FHB,with an ACA of 75% at 5 days after inoculation(~9.73% DS)and 100% at 10 days after inoculation(~18% DS).However,the combined spectral features did not significantly show better CA than the single feature.RF showed the highest CA.In addition,the multivariate Knn regressionbased model performed better(R2 = 0.86,RMSE = 10.36)than the single variable-based(WFSI1,WFSI2,WFCI1 and WFCI2)(accuracies R2 of 0.80,0.82,0.80 and 0.81,respectively,and error RMSEs of 12.75,13.56,14.17 and 13.50).This study demonstrates the feasibility of using imaging and non-imaging hyperspectral and machine learning methods for early monitoring of wheat FHB at the spike and canopy scales.The results provide theoretical and technical support for the development of a crop disease monitor that can contribute to high crop quality and yield.Additionally,the results would assist in monitoring wheat diseases ensuring the high yield and quality of agricultural products and minimizing the environmental impacts of crop protection.