Detection Of Wine Grape Classification And Tannins Content Based On Near Infrared Hyperspectral Imaging

The variety,producing area and chemical content of wine grape are the important indexes that affect the quality of wine grape,and have an important influence on the quality of wine.Wine grapes from different varieties and regions contain significantly different chemicals,so the wines they make have unique flavor and nutritional value.In recent years,near-infrared high spectral imaging technology has been widely used in wine grape internal quality detection.In this study,near-infrared hyperspectral technology,combined with stoichiometry,machine learning and deep learning methods,was used to rapidly detect the variety,origin and tannins content of wine grapes at maturity,so as to establish an effective prediction model.The results show that the NIR hyperspectral imaging technology provides a reliable tool for qualitative analysis and quantitative detection of wine grape.The main research contents and conclusions of this paper are as follows:(1)Hyperspectral imaging technology was used to establish a variety identification model for6 red and 6 white grape varieties.In the identification of grape varieties,the method of Mahalanobis distance is used to remove the abnormal spectral data from grape samples of different varieties.By comparing the SVM models of different spectral preprocessing methods,we can see that the performance of S-G smoothed model is the best.Finally,PCA load method is used to select the characteristic wavelength of spectral data and establish the classification model of different grape varieties.Compared with random forest and Ada Boost algorithm,SVM model has the best classification effect.Under the model,the average classification accuracy of the training set and the testing set of red grape varieties is 93.06%and 90.01%,respectively,and that of white grape varieties is 84.77%and 81.09%.The results show that hyperspectral imaging technology has great potential in the rapid identification of grape varieties.(2)The origin classification model of Cabernet Sauvignon was established by hyperspectral imaging technology and deep learning method.The classification effect of SVM model under different pretreatment methods was analyzed and compared,and the optimal spectral pretreatment method was determined.One dimensional convolution neural network(CNN)and residual network(Res Net)architecture are mainly designed.By comparing SVM model and two convolution networks,Res Net model has the best classification effect.The accuracy of training set and test set is 99.16%and 97.12%,respectively.At the same time,the last layer of Res Net feature data is visualized by t-SNE.The results show that spectral imaging technology combined with deep learning can effectively identify the grape producing area.(3)A detection and analysis model of tannins content in grape peel at mature stage was established.This paper introduces the results of chemical statistics and spectral pretreatment of tannins content in grape samples,puts forward a new method of wavelength selection Monte Carlo Frequency(MCF),and compares MCF with conventional wavelength selection methods spa and cars.Finally,the prediction model of grape tannins content is established by SVR.The model had the best prediction effect after MCF dimension reduction,and the prediction set R~2 and RMSE reached 0.82 and 0.09mg/g,respectively.The results show that the tannins content of wine grape is highly correlated with hyperspectral data,and the proposed MCF algorithm can effectively reduce the dimension of hyperspectral data.