Research On Detection Method Of Wheat Ear Scab Infection Based On Hyperspectral Imaging Technology

Wheat is one of the main crops in the world,and the annual total output and total trade volume for wheat are the first ind food crops.Wheat scab is a worldwide disease caused by Fusarium graminearum,which will cause to wheat ear rot and other damage.The toxin produced by wheat scrab,such as deoxynivalenol(DON)will seriously harm human health and bring great hidden trouble to the food safety.With the global climate warming and change of cultivation system and methods like straw returning to the field,wheat scab has broken out in a large scale all over the world,mainly distributed in humid and semi-wet areas.There is a growing trend for the occurrence area of wheat scab in China.In order to prevent the outbreak of wheat scab in a large area,it should be carried out disease diagnosis and prevention in the early stage.However,the detection method for wheat scab is relying on artificial discrimination at present,and the method is subjective,time-consuming and laborious,prone to miscalculation.Therefore,it is an urgent problem for wheat scientists in China and even in the world to solve the prevention,control and diagnosis.In this paper,in order to solve the diagnosis and prevention diffcult for wheat scab,hyperspectral image system wiht the visible near infrared wave band(400～1000nm)was applied for wheat scab detection.Wheat ear infected by Fusarium graminearum was taken image and collected hyperspectral image data,then the characteristic band of the reflectivity file was extracted,and image processing and deep learning algorithm was used for the characteristic band image.The clasifaction of wheat head scab will lay a foundation for accurate diagnosis and problem.The main research contents and conclusions of this paper are as follows:(1)A method for ear collection of wheat scab was designed based on hyperspectral image.Single flower inoculation method was used to artificially inoculate the wheat ears for Yangmai 23 with spore suspension during the flowering period.Samples of infected wheat ears of different disease grades were obtained by three times injections.The samples were classified and the parameters of the hyperspectral imager were adjusted to collect the hyperspectral image of the wheat ear samples,the original spectral data of the infected wheat ear was obtained,and the reflectivity data of the wheat ear region was extracted by the region of interest extraction method.(2)A method for extracting characteristic waveband based on spectral data of wheat scab was studied.The reflectivity data using different pretreatment algorithm like MSC,SNV and characteristic waveband extraction method like SPA,XLW to process the extracted reflectivity files to get different characteristic waveband combination and modeling and forecasting the reflectivity data of characteristic waveband through SVM and PLS-DA according to both the infected and normal wheat ears,choose the best characteristic waveband of 682nm,714nm and 553nm.(3)A method for testing wheat scab classification based on characteristic waveband image was proposed by image processing method.On the basis of characteristic waveband images extracted by hyperspectral technology,different coefficients were added and subtracted to obtain the gray images of wheat ear and then carried out image pre-processing and denoising.Furthermore,the image was segmened by the variance method between the largest categories and integrated the corrosion and expansion to remove the residual noise of wheat awns.To alphashape convex hull extracted contour according to the outline which fitting out the main shaft of wheat ear,through the analysis to the segmentation features of main shaft in the infected area and grain binary map to obtain the wheat features and then calculated the spike rate and ill grade,compared with artificial discriminant result,the accuracy of wheat scab in classification 1,2,3,4 were 88.8%,85.7,79.1%,87.09%respectively.The reason for the error was that the segmentation of wheat spike image was not accurate due to wheat awn and other noises,and the calculation accuracy cannot be improved due to the limitation of characteristic extraction algorithm in later period.(4)A method to identify the classification of wheat ear scab based on hyperspectral characteristic waveband image under deep learning was studied.The characteristic wavebands of 682nm,714nm and 553nm extracted by XLW algorithm were combined to generate 3-waveband images,and the combined images were flipped and the brightness was adjusted to obtain the expanded feature image data set.Adjusted the preliminary selection frame size of Faster RCNN and DSSD algorithm,using AlexNet,ResNet101 neural network model,batch number and learning rate to optimize the training parameter,eventually determined that the batch number of Faster RCNN was 64,the initial learning rate was 0.01 and after 15000 iterations,the learning rate was down again by one over ten,the learning rate of DSSD algorithm was 8,the initial learning rate was 0.01 and after 10000 iterations reduced at one over ten,so as to obtain the best training model under the best parameter combination.To compare the algorithm accuracy of different models through testing,and finally it was concluded that the accuracy of ResNet101+Faster RCNN algorithm was the highest,and the tested accuracy of deep-learning for the different classifications of 1,2,3 and 4 were 100%,94.3%,91.6%and 100%,respectively.