Study On Rapid And Nondestructive Detection Method Of Mildew Wheat Seeds Based On Hyperspectral Imaging Technology

As one of the three major food crops in China,wheat seeds are very easy to cause mildew suffering from high temperature and high humidity during the harvest season every year,and poor airtightness during storage process.Wheat mildew will produce toxins,which is harmful to human beings,endangers human life and health and causes great economic losses.Usually detection of mildew is required before wheat seeds are processed,however,the traditional detection methods are too complicated and cannot meet the industrial online detecting problems.Therefore,it is particularly important to develop accurate,rapid and nondestructive methods for the detection of mildew infection in wheat seeds.In this study,the growth time model of wheat seeds infected with mildew was constructed based on hyperspectral imaging technology combined with chemometric analysis.And the predictive regression of mildew spore content in wheat seeds was analyzed and the degree of mildew in wheat seeds was visualized.The specific researches are as follows:Firstly,hyperspectral imaging system with wavelength of 900～2500 nm was used to discriminate the growth time of Aspergillus ochraceus infected wheat seeds.Hyperspectral images of all wheat seeds were acquired and average spectral data within the sensibility region were extracted.The acquired data were pre-processed by three methods:Savitzky-Golay(SG)filter smoothing,1st derivative(1d)differentiation,and standardized normal variate(SNV).Least angle regression(LAR)and successive projections algorithm(SPA)were used to select the feature wavelengths.Support vector machine(SVM)and partial least squares discrimination analysis(PLS-DA)classification models were built based on the feature bands and the pre-processed full bands.The six models being comprehensively compared,the best wheat seed inoculation Aspergillus ochraceus growth time discrimination model was established based on the SNV+SG-SPA-SVM algorithm,with accuracy of 96.39%for the training set and accuracy of 94.64%for the test set,respectively.Secondly,based on a hyperspectral imaging system in the wavelength range of 900～2500nm,a regression model was developed using quantitative analysis to predict the spore content in Aspergillus ochraceus infected wheat seeds.Hyperspectral images of all wheat seeds were acquired and average spectral data within the sensitized area were extracted,thus determining the spore content according to GB 4789.2-2022"Determination of total bacterial colony of food microbiological test".LAR and SPA were used to select the characteristic wavelengths.Based on the feature bands and the full bands,support vector regression(SVR)and partial least squares regression(PLSR)classification models were built.The six models being comprehensively compared,based on the SPA-SVR algorithm,the best determination model of the spore content of wheat seeds inoculated with Aspergillus ochraceus was established with coefficient of determination(R_p~2)of 0.9681,root mean squared error of prediction(RMSEP)of 0.3926,and residual predictive deviation(RPD)of 5.1162 for the prediction set,respectively.Finally,in order to better distinguish the degree of mildew,according to GB 4789.2-2022"food microbiological inspection of the determination of the total number of colonies"to determine the spore content,four levels of mildew degree of wheat seeds(safe,critical,hazardous and serious hazard)were established.Visualization of different levels of Aspergillus niger infection in wheat seeds was conducted using a hyperspectral imaging system in the range of 900～2500 nm.Hyperspectral images of all wheat seeds were acquired and average spectral data within the sensibility region were extracted.Feature wavelengths were selected by binary fireworks algorithm(BFWA),competitive adaptive reweighted sampling(CARS),random frog(RForg)algorithm and shuffled frog leaping algorithm(SFLA),respectively.Based on SVM,K-Nearest Neighbor(KNN),Le Net-5 and VGG-16 algorithms in convolutional neural networks(CNN),the hyperspectral monochromatic images corresponding to the characteristic wavelengths were used to construct pixel recognition models.After a comprehensive comparison of the four models,it could be seen that when monochromatic images of four components(1032 nm,1452 nm,1126 nm and 1901 nm)were combined under the Le Net-5 algorithm,the accuracy of the training and prediction sets for recognition was 100%and 99.86%,respectively.And area under curve(AUC)in the receiver operating characteristic(ROC)curve of the prediction set reached 0.9969.The above study provides a theoretical basis for the construction of a complete method for the nondestructive and rapid detection of the degree of mildew in wheat seeds.