Study Of The Early Detection Of Fungal Disease Of Barley Seedlings Based On Hvperspectral Imaging Technology

This study targets on the early detection of fungal disease of barley seedlings based on hyperspectral imaging technology.Firstly,based on the full-band and characteristic bands of visible/near-infrared spectra,a quantitative determination model of metabolic components was built.the classification and identification of disease spot in different periods of infection based on spectral features and image texture.Secondly,we focus on the spectral features and image texture features of disease spot in different periods of fungal disease.We explored the potential of using the pectral features and image texture features of hyperspectral images to identify the different periods of infection.The main achieves were reached as follows:(I)Three species of barley seedlings with different resistance to Magnaporthe oryzae were selected as experiment subjects.It is found that the metabolic components,including chlorophyll a,chlorophyll b,ascorbic acid and malondialdehyde,were significantly associated with the growth state and the period of infection of barely.Several chemometric techniques were used to establish quantitive determination models of the metabolic components of barley based on Hyperstpectal imaging.Three characteristic wavelength selection algorithms of the interval least squares algorithm(biPLS),Successive projection algorithm(SPA)and competitive adaptive re-weighted sampling(CARS)was compared.12-27 characteristic wavelengths were selected and were used to establish PLS and LS-SVM regression models.The coefficient determination value Rv2 is higher than 0.800 for training set.The coefficient determiantion value Rv2 reached above 0.710 for all three components except for chlorophyll b.(2)The hyperspectral images of three speceise of barley seedling were collected.A discrimination model of four infection periods of barley were build based on the average spectra.We established and evaluated four classification models including L1 Logistic regreesion,L2 Logistic regression,Linear SVM and Linear discriminant analysis(LDA)were build based on the selected characteristic wavelengths.The optimal model is the combination of 37 characteristic wavelengths selected by CARS and LDA classification method.The accuracy for training set and testing set are 96.0%and 98.4%,respectively.This model reduces the computational complexity of the model,and successfully obtains the effective information in the hyperspectral data as well as establishes a highly accurate discrimination model.(3)The hyperspectral images of three speceise of in vivo barley seedling in 4 different infection periods were collected.We study the mapping of score of the first 3 principle components of the hyperspectral images base on principal component analysis(PCA).Identically the mapping of abundance of the 4 end-members of the hyperspectral images base on Vertex component analysis(VCA)were also extracted.The visualization and monitor of infection,incubation and out-breaking periods of the fungal disease were realized.(4)Texture features of hyperspectral images of the barley from different infection period were studied,including 0,48-72 h and 96-120 h after infection.In terms of no texture imaging,the 3D hyperspectral images first were projected into the 2D space of principle components(PCs)with the help of PCA,and then the resultant PCs are processed by GLCM and GLRLM to generate the texture features.The abundance of 4 end-members of VCA were also used as the 2D grayscale images.Three classification models of the infection period of barley were established based on the extracted texture features.Finally,the optimal model reached the accuracy of 93.02%for training set and of 96.92%for testing set.