Research On Multi-cycle Extraction And Combination Identification Of Hyperspectral Information Of Fusarium Head Blight In Wheat

Wheat is one of the most important food crops in the world and has a wide planting area in China.Fusarium head blight(FHB)is a major disease of wheat,which has strong infectivity.In addition,it will not only lead to wheat yield reduction or even crop failure,seriously affect the quality of wheat seeds,but also secrete a large number of toxins which seriously endanger human and animal health.Therefore,the identification of FHB in wheat is of great significance.In this thesis,hyperspectral information extraction technology and identification model of FHB in wheat grains were taken as the starting point,and image processing technology,feature extraction technology and classification algorithm were used as means to confirm the feasibility of hyperspectral imaging technology for rapid identification of FHB in wheat.At the same time,the optimization of feature information extraction method and identification model is studied,and the multi-cycle information extraction method and combined identification model are constructed,which improved the accuracy of feature information extraction and identification,reduced the rate of misjudgment and missed detection,and constructed an efficient,accurate and visual identification model for wheat grains caused by FHB in wheat.The main contents and results of this thesis are as follows:(1)Research on hyperspectral image preprocessing technology of FHB in wheat grainsIn this thesis,hyperspectral images of healthy wheat and FHB in wheat provided by the Second Granary Program Normal Farm of the Chinese Academy of Sciences in visible to near infrared(470-1100 nm)bands were collected by hyperspectral imaging system.In the aspect of spatial image,gray variance method is used to select the optimal band for image processing as the evaluation index of clarity.Then,the maximum interspecific variance method is used to extract wheat mask combined with gray linear stretching and opening operation,and principal component analysis is used to enhance the accuracy of mask extraction.In spectral pretreatment,Savitzky-Golay(SG)convolution smoothing algorithm,multiplicative scatter correction and standard normalized variate were used to process wheat spectral data.Finally,SG convolution smoothing was used for spectral pretreatment.In the aspect of hyperspectral feature extraction,principal component analysis(PCA)and successive projections algorithm(SPA)are used to reduce the dimension of hyperspectral data.After PCA transformation,six principal components contain more than 99% of the original information.SPA algorithm extracts eight characteristic wavelengths,which are 479.9,498.2,543.8,630.5,726.2,828.9,904.1,920.1nm.Effective data preprocessing and feature extraction provide a guarantee for the establishment of efficient and fast identification model of FHB in wheat.(2)Extraction and optimization of FHB in wheat characteristic informationIn the process of extracting the spectrum of wheat samples,a method based on K-means clustering algorithm combined with multiple cycles of kappa coefficients is proposed to extract the optimal training samples.Finally,500 sample spectra of each class are used as training set,and 10 000 sample spectra of each class are used as test set.The model is evaluated by combining the total classification accuracy and kappa coefficient.In PCA and SPA feature spaces,several identification models of FHB wheat were constructed by using five classification algorithms: Spectral Angle Matching,K-means clustering,k-Nearest Neighbor,Linear Discriminant Analysis and Support Vector Machine(SVM).The results show that the classification model constructed by SVM in SPA feature space has the best classification performance.The classification accuracy of training set is 91.1%,test set is 88.84%,and kappa coefficient is 0.7767.(3)Establishment and validation of Hyperspectral Feature Combination recognition model for FHB in wheatOn the basis of SVM model for FHB in wheat identification,two methods to further improve the identification accuracy of FHB in wheat were proposed in this paper: MSC-based quadratic SVM identification method and comprehensive PCA and SPA feature space SVM identification method.The former method uses SVM model and SVM(MSC)model to improve the classification accuracy of test set in SPA feature space to 88.98%.The latter method synthesizes the information of samples in two feature spaces,and chooses a more suitable classification model in different reflectivity intervals,which improves the classification accuracy of the test set to 90.18%.The hyperspectral images of wheat grains mixed with 50% FHB infection rate were collected as test samples with unknown labels in the future,and SG-SVM(PCA + SPA)model is used to visualize the original location of FHB in wheat.The above research results show that hyperspectral imaging technology combined with data processing algorithm can realize the original location,fast and visual identification of FHB in wheat,help to improve the detection efficiency,detection quantity,reduce the rate of missed detection,provide guarantee for wheat storage,transportation and processing,and ensure the safety of wheat practically.This study provides a new research idea for FHB in wheat detection,lays the algorithm foundation for the development of automatic detection equipment for FHB in wheat,and has broad application prospects.