| Rice is one of the main crops in China,and the nutritional components of different varieties of rice seeds are quite different.In addition,seed vigor,as an important index to evaluate rice seed,has a great impact on the yield of rice.Therefore,it is of great significance for agricultural production to select excellent varieties of rice seeds with high vigor.Traditional seed quality detection methods(such as tetrazole method,enzyme activity determination method,germination experiment method and so on)are difficult to be vigorously popularized in agricultural production due to their disadvantages of time-consuming,laborious and high cost.Hyperspectral imaging technology has the advantages of fast,non-destructive and high-precision testing,and has been widely used in the field of agricultural product quality detection.Therefore,in this paper,five varieties of rice seeds were taken as the research object,and hyperspectral imaging technology was used to carry out the identification of varieties.And on this basis,the vigor level testing of rice seeds was studied for a single variety.The main contents and conclusionsare as follows:(1)The average spectral curves in the visible near infrared(Vis/NIR)range of five rice seed varieties were analyzed,and the response range of seed pigment and organic matter content in the spectral range was discussed.The feasibility of using Vis/NIR hyperspectral imaging technology to identify seed varieties was clarified.In addition,the average spectral curves of three vigor levels of single variety in the near infrared(NIR)range were analyzed in detail,the correlation between the changes of organic matter content and cell structure in seeds and NIR spectra was illustrated,and the feasibility of using NIR hyperspectral imaging technology for seedvigortesting wasexplored.(2)The hyperspectral images of five rice seed varieties(Shengdao 22,Xudao 9,Shaonuo9714,Baohan 1and Lvhan 1)in the Vis/NIR(400nm-1000nm)range were obtained,spectral and image information were extracted.Firstly,standard normal variable transformation(SNV)combined with de-trending(DT)was used to preprocess the spectral information.Then,variable iterative space shrinkage approach(VISSA)and bootstrapping soft shrinkage(BOSS)were used to select feature variables of spectral and image information.After that,the support vector machine(SVM)models based on spectral and image features were constructed,respectively.Afterwards,the performance of the models was analyzed and the best spectral and image feature variables were determined.Due to the limitation of single feature model,SVM model based on fused feature was established,and grey wolf optimizer(GWO)and improved grey wolf optimizer(DE-GWO)were used to optimize the parameters of SVM model.The results showed that the model based on fused feature had higher accuracy,and the accuracy reached 96.30%.In addition,the accuracy of DE-GWO-SVM model was higher than GWO-SVM model,and the accuracy of DE-GWO-SVM model reached 98.52%.From the above analysis,it could be seen that feature fusion could improve the classification accuracy of the model.Compared with GWO algorithm,DE-GWO had more superiority in model parameter optimization.Therefore,the DE-GWO-SVM model based on fused feature had excellent classification performance in riceseed variety identification.(3)For a single variety,three vigor levels of rice seeds(high vigor,low vigor and no vigor)were selected as the research objects,and the hyperspectral images in the NIR(871nm-1766nm)range were obtained.From which,the spectral data were obtained.Firstly,wavelet threshold denoising(WTD)combined with first derivative(FD)was used to preprocess the obtained spectral data,and principal component analysis(PCA)and stacked auto encoder(SAE)were used to extract the feature variables from the preprocessed spectra,respectively.Then,the SVM classification model was constructed by using the feature variables obtained by PCA and SAE,and the model with higher classification accuracy was selected.Finally,the parameters of the selected model were optimized by GWO and DE-GWO.The results showed that the accuracy of SVM model based on SAE deep feature was higher,which reached 96.25%.The performance of SVM model was further improved after GWO and DE-GWO optimization.The accuracy of DE-GWO-SVM model was higher than GWO-SVM model,and the accuracy reached 99.25%.Therefore,hyperspectral imaging technology combined with deep learning algorithmcould achieveaccurate classification ofriceseed vigorlevels.In this paper,hyperspectral imaging technology combined with machine learning algorithm was used to achieve high-precision variety identification.On this basis,the seed vigor of a single variety was accurately detected.The research results provided an effective method for selecting high-quality variety with high vigor,and laid a theoretical foundation for the production of relateddetectioninstrument. |
