Study Of The Early Detection And Visualization Of Barley Disease Based On Spectral Imaging Technology

Barley(Hordeum vulgare L.)is one of the major cereal crops,which plays an important role in global food production.Barley cultivation is widely distributed,so that it would be infected by a large number of different kinds of diseases,which leads to the reduction of barley production and economic losses in agricultural production.Therefore,it is of great significance to carry out rapid detection and research on a variety of barley diseases and contribute to the early and accurate prevention of disease.This study mainly applied hyperspectral imaging(HSI)and also Raman spectroscopy technology,to carry out early detection of barley samples under stress of three fungal diseases.Based on HSI,this study realized the diagnosis of the disease stage of Magnaporthe oryzae(M.oryzae)infected barley,proposed a disease identification method that can pre-visualize disease spots;realized the rapid prediction of the severity of powdery mildew infection,and combined with spectral and disease-spot texture information to predict the future development of diseases;realized the identification and analysis of barley under the infection of powdery mildew from canopy scale;realized quantitative detection of various indicators such as pigment and antioxidant composition of barley leaves under the infection of spot blotch.Based on Raman spectroscopy and displacement platform,this study realized the macro-scale Raman imaging,established a linear prediction model of pigment content,and further fulfilled the pigment visualization of leaves under the stress of powdery mildew based on the Raman‘fingerprint'peaks.The detailed conclusions are as follows:(1)For the purpose of symptom identification and detecting disease development process of M.oryzae infected barley.This study applied HSI to collect sample data of different time periods after infection and used CARS algorithm to analysis the average spectra and extract 30characteristic wavelengths related to pigment and tissue structure.After that a better classification model was obtained in combination with the LDA method,and the classification accuracy of the calibration set and prediction set was 100 and 98.1%,respectively,which realized the early detection of the disease stage.It is found that the application of spectral unmixing based on full-spectrum analysis can be used on hyperspectral images analysis to identify the early lesions more accurately.Using VCA to extract spectral features,then the abundance map can be obtained by spectral unmixing analysis,which can clearly show the disease area at 24 hours after the infection,even before the recognization of naked eye from RGB image.At the same time,this method also clearly shows the complex distribution of diseases on the leaves in the late period of infection..(2)Based on the characteristics of powdery mildew infection,the proportion of powdery mildew disease spots can be used to quantify the degree of stress on barley leaves.This study analyzed the hyperspectral images of powdery mildew infected barley leaves.A linear model between the average spectra of samples and the proportion of infection was established,the R~2 of calibration set was 0.858,and the R~2 of the prediction set was 0.821,indicating that the average spectrum could well reflect the proportion of infection of powdery mildew.Further study combined the average spectral information and the texture information of the disease spot from current sample,to predict the proportion of powdery mildew infection in the next 24 hours.7variables were selected to obtain the optimal quantitative model,R~2 of calibration set was 0.87,and the R~2 of the prediction set was 0.80,which enables the prediction of future disease development.The selected seven variables are:sum of squares,maximum probability,contrast,homogeneity,inverse different moment from statistical result of GLCM,reflectance value at679.4nm and the last class of LBP statistics.(3)Based on HSI,the visualization of early infestation sites of barley under stress of powdery mildew is studied from the canopy scale.This study found that because of the different depth of the canopy layer,the spectral correction with whiteboard reflectance selected from a certain location,will have a great impact on the data,while the impact of camera focus is relatively limited.In addition,larger angle of the incident light would make the collected information low dispersion.After using 0-1 normalization,the faded green area of leaf representing aging is divided by K-means.Then VCA is used combined with FCLS to further distinguish the diseased region and the healthy area.By comparing with artificial identification results,this recognition method obatained error less than 10%late in the stress period.(4)Using hyperspectral imaging,the changes of three pigment contents chlorophyll-a,chlorophyll-b and carotenoids,three antioxidant composition ascorbic acid,glutathione and malonaldehyde under the stress of spot blotch,were analyzed.The quantitative model of each indicator is established based on the average spectrum,and the characteristic wavelengths of each indicator are selected,respectively.The results of all models established for all indicators are good with R~2of calibration set higher than 0.91,and R~2of prediction set higher than 0.84.The model can be used well for the prediction of the chemical value of the sample,but in the visualization of the composition distribution,it is difficult to predict the content of all pixels effectively because the average spectrum actually masks the huge spectra difference between the pixels of the lesion area,the faded green area,and the healthy area.(5)The combination of portable Raman spectrometer and high-precision displacement platform to obtain macro-scale spatial Raman information of barley leaves is realized.In the study,the pigment content of the sample was studied under the influence of powdery mildew,and the"fingerprint"peak of the pigment was selected from the Raman spectrum after SWi MA processing,and a linear quantitative model of the pigment content was established.Combined with Raman mapping,the visualization of the three pigment distributions of barley leaves is realized,and the mean value of the pixel prediction is within 8%offset from the actual pigment value of the sample.In addition,for barley samples infected by spot blotch,the R~2of prediction set of all three pigments were higher than 0.84,and the slope of the scatter distribution fitting line was in the range of 1±0.13,indicating that the deviation between the real value and the predicted value was very small,which also proved the effectiveness of the pigment quantitative model based on the"fingerprint"peak of Raman spectroscopy.