Identification Methods Of Fusariumhead Blight Based On Spectral And Image Feature Information

Fusarium head blight(FHB)is a major disease in wheat production,and always causes the loss of wheat yield and wheat quality.This disease is evaluated by certain expert experience of experienced agronomists or technicians,which has the shortages with manpower consuming and low efficiency,is difficult to adapt to the real-time and accurate requirements of disease control currently.Therefore,it is very important to study the rapid diagnosis methods of FHB severity during the growth period of wheat.Hyperspectral imaging,with the data advantage of interatingspectral and image information,has the technical advantage in the rapid detection of crop diseases.Up to date,there is few effective method for the rapid detection of fusarium head blight.This paper attempts to combine image processing,spectral index construction,machine learning and deep learning technology to study methods on the diagnosis of fusarium head blight,with a view to providing technical support for the prevention and treatment of FHB spraying and the assessment of post-disaster losses.The main research contents are as follows:(1)Construct a new Fusarium Disease Index(FDI)A new spectral index FDI that suitable for key growth stages was proposed based on the analysis of spectral difference of infected wheat ear in the range of 374~1000nm.Random forest(RF)algorithm was used to select the sensitive bands at the flowering period,pustulation period and the two growth periods,which have the highest positive weighting coefficient and the lowest negative weighting coefficient,and then the FDI index was constructed by normalization of the screened sensitive bands and wavelength differential.The research found that,(1)the sensitive bands of wheat FHB in different growth stages are located in the range of 560~6809)9)9)9).The FDI index was calculated by the identified sensitive bands,and the linear regression models were established between FDI index and the severity of wheat FHB in different growth periods,the coefficient of determination(R~2)are higher than the common 17 disease spectral indexes.(2)The predicted results at the flowering and pustulation period(R2 of 0.94 and 0.96,respectively)are better than those in the comprehensive growth stage(R~2 of 0.90),and the prediction effects of test set for comprehensive growth stage on the single growth period(R2 of0.82 and 0.94,respectively)are lower than that calculated from the single growth period.(2)Double-side feature extraction of wheat ear and identification of severity of fusarium head blight.Due to the three-dimensional structure of wheat ear,the spectral and image features of wheat ear with the front and back sides(namely A and B)were extracted for recognition of disease severity.Firstly,the PCA algorithm is applied to select the single-band image with principal component.The image texture features such as mean value,standard deviation,entropy,energy and so on were extracted using double tree complex wavelet algorithm and gray co-occurrence matrix.The features of RGB and YDbDr color spaces(including first order moment,second order moment and third order moment)were extracted,respectively.Secondly,the GBDT algorithm was used to screen the weights of the image and spectral features,and the method of sequential backward selection(SBS)was adopted to determine the optimal features of image and spectra.Regression model was developed by RF algorithm with different features.The study results showed that the models were established by spectral and image features of A and B sides for wheat ear,there is an increase change of the determination coefficients of the model forspectral,image features,and the fusion features of image and spectra of wheat ear.In particular,the integrated features of image and spectra,which has the highest R~2 and the lowest RMSE of 0.89 and 6.52,respectively.Finally,a lightweight deep learning network model was established to explore deep learning method to predict the recognition effect of the wheat FHB,with the best performance of the fusion of double-side image and spectral features.The study found that compared to the models established by partial least squares regression and support vector regression,the lightweight convolutional neural network model has the highest determination coefficient(R~2=0.97)and the lowest root mean square error(RMSE=3.78),.In summary,the developed FDI disease index and the lightweight convolutional neural network model constructed in this paper can provide method support for the rapid and accurate detection of wheat fusarium head blight,will have a good application prospect.