Research On Wheat Based On Hyperspectral Imaging Technology

During the harvest period,wheat was highly susceptible to spike sprouting due to rainy weather,which could seriously cause large yield reduction.Untimely drying or poor management during storage could also directly lead to normal wheat sprouting,causing a series of changes in wheat quality.Too much sprouted wheat could cause economic losses when harvesting.At this stage,wheat germination was mostly judged by visual observation,but it was not easy for the naked eye to detect the subtle changes in the appearance of wheat at the beginning of germination.If determined by chemical methods,it would take a long time,the steps were tedious and destructive to the sample.Therefore,it was necessary to find a fast and nondestructive method to detect the degree of wheat germination as well as to monitor and warn the germination of wheat.In this paper,hyperspectral imaging was used to detect the germination status of wheat.Firstly,six different states of wheat were obtained by germination test:normal wheat,bulging stage,skin cracking stage,dewy,one-half of the seed length,and seed length equal to the seed length,and their spectral information and image information were collected by hyperspectral.On the one hand,by processing the spectral information(denoising,extracting characteristic wavelengths)and establishing different models,the internal chemical information reflected by the spectra was used to discriminate the degree of wheat germination;on the other hand,the external information(grain shape,texture)was used to discriminate the wheat germination status by collecting the image information.The number of wheat in each germination state and the percentage of sprouted wheat were counted and the results were visualized.Finally,the regression model ofα-amylase activity of germinating wheat was established by fusing the two kinds of information to judge the germination status from the perspective of amylase activity.The contents and conclusions of the study were as follows.(1)Spectral information from hyperspectral techniques(850 nm-1700 nm)was used to classify and identify the germination status of wheat.Firstly,the Monte Carlo cross-validation method eliminated abnormal samples and the K-S method divided the sample set.Two models,PLSR and LS-SVM,were developed to classify and identify germinated wheat using full-band spectral information.The comparison found that the prediction set recognition rate of LS-SVM was higher at 89.52%.Then,LS-SVM models were established by eight spectral preprocessing methods,including S-G(5 points),S-G(7points),S-G(5 points)first-order derivative,S-G(7 points)first-order derivative,S-G(5points)second-order derivative,S-G(7 points)second-order derivative,MSC,and SNV,and the best preprocessing method was determined to be SNV,with a classification accuracy of91.70%.After optimizing the regularization parameters of the LS-SVM model with two parameters of the kernel function by PSO,the recognition accuracy of the optimized model was improved to 93.14%.Next,49 feature wavelengths were selected by the competitive adaptive reweighting algorithm(CARS)to build the SNV-CARS-PSO-LS-SVM model with an accuracy of 94.13%and 99.78%for germinating wheat.Finally,the SNV-CARS-PSO-LS-SVM model was identified as the optimal model for wheat germination classification and the results were statistically and visualized.(2)The hyperspectral image information was used to classify and identify the germination status of wheat.Firstly,seven morphological information of different germination states of wheat were extracted,and then four texture feature extraction methods,namely grayscale-gradient coeval matrix method,local binary method,differential statistics method,and Tamura method,were used to obtain texture features,and morphological features were fused with texture features extracted by the four methods using normalization,respectively.A classification recognition model of PSO-LS-SVM as well as PLSR for wheat germination state was established.The texture extraction method with the highest prediction accuracy was the grayscale-gradient covariance matrix method with an accuracy of 57.67%for LS-SVM.37.67%for PLSR.The prediction was not satisfactory compared to the spectral information.The visualization process changed the six classifications to three classifications(skin cracking stage and later classified as sprouted wheat)based on the recognition results of the model on the training set,and the model accuracy improved to81.16%,and the recognition accuracy of sprouted wheat was 80.63%.And the classification results were visualized.(3)The regression prediction model ofα-amylase activity based on spectral information was established,and the coefficient of determination R~2of LS-SVM model at characteristic wavelength was 0.66 and that of PLSR was 0.61.The prediction model of enzyme activity based on image information was finally established by fusing the spectral information of germinated with image information to establish a quantitative model ofα-amylase activity.The prediction model established by image information LS-SVM model had R~2of 0.5 and PLSR of 0.33.The regression model effect was very unsatisfactory and finally the image information of germinated wheat was fused with spectral information by normalization for a total of 45 features.The R~2of the coefficient of determination of the model built by comparing the single use of image information and spectral information were both slightly improved,and the R~2of the LS-SVM model was 0.71 with PLSR of 0.65.The model fit was not high enough.Also,the recognition rate for classifying wheat based on enzyme activity was not high,but there was 96.43%recognition rate for wheat with shoot length as seed length.Follow-up studies may consider the effect of wheat species as well as obtaining more enzyme activity data of wheat to supplement the model,while the enzyme activity thresholds for each state of germinating wheat are to be studied.