Research On Key Techniques Of Green Tea Classification Based On Aroma And Near Infrared Spectroscopy Features

Tea is an important economic crop in our country,which occupies a large share of exports.However,the existing classification of tea categories is mostly subjectively evaluated by people,and the system is relatively chaotic,which is not conducive to the promotion of the industry.Therefore,this article takes green tea as the research object,establishes a classification model based on aroma and near-infrared spectroscopy data,and conducts research on the problems of low classification accuracy using aroma data,high collection cost of spectral data,and high dimensions of spectral data in the available technology.A green tea classification method using optimized aroma characteristics that filitered by information gain is proposed.Firstly de-baseline adjusting and Savitzky-Golay filtering method is used to reduce noise interference of aroma data.Aiming at the problem of insufficient or redundant information caused by using one or more feature of aroma data,such as mean and variance in the existing classification models,a feature ranking method based on information gain is proposed,and the optimal feature set is selected according to the ranking,and inputted into the LDA model for classification,which has improved the distinguishability of feature sets and the accuracy of green tea classification.A green tea classification method based on Bagging ensemble learning is proposed,which using a small range of wavelength spectrum.Aiming at the problems of high data collection costs caused by using a wide range of wavelength data,and the low accuracy of classification using a small range of wavelengths,an ensemble learning classification model based on the Bagging method is proposed.First,the standard normal variable transformation method is used to reduce the scattering interference during data acquisition;secondly,the decision tree,K nearest neighbor method,and linear discriminant analysis are used as the basic classifiers;then,based on the Bootstrap sampling method,the Bagging ensemble learning model is established to classify green tea.Experiments confirm that this method can improve the discrimination accuracy of the classification model which using small-range wavelength data effectively.Joint classification methods of aroma and spectrum are proposed,in decision-level,we use probability data and perceptron model;in feature-level,we use data/feature fusion optimization.In the decision level,in view of the problem that the simple vote of aroma and spectrum classification results is hard to classify the green tea accurately,it is proposed to use the classification probability of respective classification models of the aroma and spectrum as the fusion data of the decision level,and use the perceptron to classify.In the feature level,aiming at the information redundancy problem of aroma features/spectral data,three feature combination and selection methods is proposed: directly combine aroma features and spectral data,combine aroma features optimized based on information entropy gain with original spectral data,and optimize aroma and spectral features/data based on information entropy gain.And finally use the linear discriminant analysis method to classify the green tea.