Research On Monitoring Method Of Winter Wheat FHB Based On Near-Earth Hyperspectral Technology In Anhui Province

In this thesis,near-earth non-imaging hyperspectral data was used to study the detection methods of wheat Fusarium head blight(FHB)on different scales.Taking wheat FHB as the research object,sensitive features variables were extracted at the wheat ear scale and canopy scale,respectively.And different classification algorithms were used to establish a monitoring model to monitor the severity of wheat FHB in the field.The study results can provide guidance for agricultural departments and managers on the monitoring of wheat FHB,disease control and field management.The main research work is as follows:(1)On the scale of wheat ears,the wheat FHB was identified from different measurement angles.Non-imaging hyperspectral data were first processed by spectral standardisation.Spectral features of the first-order derivative,the spectral absorption features of the continuum removal,and the vegetation indices were used to evaluate the ability of identifying FHB.Then,the spectral feature sets which were sensitive to FHB and have significant differences between classes were extracted from the front,side and erect angles of winter wheat ear,respectively.Finally,Fisher's linear discriminant analysis(FLDA)for dimensionality reduction,and support vector machine(SVM)are used to construct an effective identification model for disease severity at front,side and erect angles.Among selected features,the first-order derivative features of SDg/SDb and(SDg-SDb)/(SDg+SDb)that are most dominant in the model produced.The results show that: the selected spectral features have great potential in detecting ears infected with FHB;LDA combined with SVM can effectively improve the overall accuracy of the model at three angles,and the over accuracy of the side(88.6%)is better than the front(85.7%),while the over accuracy of the erect angle is the lowest(68.6%).(2)On the scale of wheat ears,this thesis proposes a new method for assessing the severity of the disease for further research on the identification of wheat FHB,and use the sample data from erect measurement angle.The original spectrum was subjected to continuous wavelet transform,combined with correlation analysis and independent sample T test to select wavelet features variables suitable for identifying FHB,and SVM algorithm wasused to construct a FHB identification model based on the new disease severity.Features that are sensitive to FHB and have significant inter-class differences were selected,including the two feature bands of the original spectrum at 491 nm and 699 nm,and six wavelet features after continuous wavelet transform of the original spectrum.The results show that:the accuracy of the model based on wavelet features is higher than that of the original spectral band features.The combination of the two features can further improve the accuracy of the model;the accuracy of models based on the new severity is higher than the models based on the initial severity.The identification accuracy of the three classifications is up to 75.56%,and the identification accuracy of the second classification is up to 91.11%;the new severity can better express the true condition of wheat FHB than the initial severity.(3)On the canopy scale,the canopy spectrum data of wheat FHB-infected samples were measured in the field,and features variables suitable for canopy scale disease monitoring were selected and a monitoring model was established.Firstly,the four traditional features commonly used in disease identification and monitoring,which are the original spectral band features,first-order derivative features,continuum removal features,and vegetation indices features,and the wavelet features obtained by continuous wavelet transform were used.Correlation analysis and independent sample T test were used to select features that are sensitive to FHB and have significant inter-class differences,the features including the original spectral feature band 635 nm,first-order differential features band 550 nm and band666 nm,Dy(the largest first-order differential value of the yellow edge),SDy / SDb(the ratio of the first-order differential sum of the yellow edge and the sum of the blue edge),vegetation index PRI,GARI,and three wavelet features F1,F2,and F3.In order to reduce the information redundancy between features,the selected optimal feature variables are reduced by LDA method as input variables,and models are constructed using Adboost algorithm and SVM algorithm respectively.The classification accuracy of the models is 70.59% and 67.65,respectively.And the accuracy of the model after adding wavelet variables is 79.41% and82.35%,respectively.The accuracy of the model after adding wavelet features is significantly higher than the models based on traditional features,and the accuracy of the Adboost model and SVM model is increased by 8.82% and 14.7% respectively.Among them,SVM model is better than Adboost model in identification of winter wheat FHB at canopy scale.