Classification Of Oolong Tea And Detection Of Tieguanyin Adulteration Based On Fluorescence Hyperspectral Imaging

Tieguanyin,as a kind of oolong tea,is the best oolong tea and is highly loved by consumers.With people’s love for famous tea rising day by day,the demand for high quality tea exceeds the supply,while the market is full of tea fraud,including replacing Tieguanyin with low-priced tea,or mixing low-priced tea in Tieguanyin,it is difficult for consumers to distinguish the authenticity of tea with the naked eye,they often buy low quality tea at high prices,seriously affecting the order of the tea market.In this study,four types of oolong tea and adulterated Tieguanyin were used as the research subjects.Fluorescence hyperspectral imaging technology was used for rapid non-destructive detection of oolong tea types and rapid detection of Tieguanyin adulteration,and optimization algorithms were used to improve the model’s prediction performance.The main work was as follows:(1)The study used fluorescence hyperspectral imaging technology to achieve rapid non-destructive detection of oolong tea classification.In this study,the fluorescence hyperspectral imaging technique was used to extract the spectral data,the raw spectra were pre-processed using standard normal variate(SNV)+first-order derivative(1st Der)and multiple scattering correction(MSC)+1st Der.Two traditional feature selection methods,principal component analysis(PCA)and variable importance projection algorithm(VIP),and a model-based feature selection method,recursive method of optimal feature elimination(RFECV),were used to feature selection.Finally,qualitative models for oolong tea classification were constructed by random forest(RF),adaptive boosting(Ada Boost)and extreme boosting tree(XGBoost).The results showed that the MSC+1st Der-VIP-Adaboost model could achieve 100%accuracy,precision and recall under traditional feature selection methods,and took the least time,only 0.111s;under model-based feature selection,MSC+1st Der-RFECV-Adaboost was the best model,with 100%accuracy,precision and recall,only took 0.389s.(2)The study used fluorescence hyperspectral imaging technology to achieve rapid detection of adulteration of Tieguanyin.SNV+1st Der and MSC+1st Der were used to pre-process the spectra,and VIP,PCA and RFECV were used to select features.The results showed that after spectral pre-processing,RF and XGBoost were able to accurately complete the qualitative identification of Tieguanyin adulterated types,both within 1s of the run time.The best model for the quantitative detection of adulteration levels was both MSC+1st Der-VIP-Adaboost,with RMSEP on the test set were 0.0000,0.0000 and 0.0308,and R_p~2on the test set were 1.0000,1.0000 and 0.9939,respectively,all with running times in the range of0.008-0.009s.The best models for detecting the level of adulteration of Tieguanyin adulterated with Benshan,Maoxie and Huangjingui after using RFECV feature selection were SNV+1st Der-RFECV-Adaboost,MSC+1st Der-RFECV-Adaboost and MSC+1st Der-RFECV-Adaboost respectively,with the RMSEP on the test set were 0.0000,0.0235 and0.0255,respectively,and R_p~2on the test set were 1.0000,0.9954 and 0.9958,with run times of 1.663s,4.144s and 2.278s,respectively.(3)The study used bayesian hyperparameter optimization algorithm to optimize the parameters in the ensemble learning algorithm to improve the model predictions.For oolong tea classification,RAW-PCA,MSC+1st Der-PCA,SNV+1st Der-PCA and SNV+1st Der-RFECV improved the accuracy after optimization by 7.04%,11.27%,7.04%and 0.69%respectively under the RF model;RAW-PCA,MSC+1st Der-PCA SNV+1st Der-VIP and SNV+1st Der-PCA were optimized to improve accuracy by 9.37%,7.04%,1.41%and 5.63%,respectively,under the Ada Boost model;RAW-PCA,MSC+1st Der-PCA,SNV+1st Der-PCA,RAW-RFECV and SNV+1st Der-RFECV improved in accuracy by 5.63%,9.86%,5.63%,6.94%and 2.78%,respectively,after optimization under the XGBoost model.In Tieguanyin adulteration detection,the model prediction ability of the spectral data was significantly improved after PCA and then quantitative detection models were built,and the predictive ability of the models were all significantly improved,with R~2 close to 1 and RMSE close to 0.In addition,among the quantitative models of adulterated Tieguanyin based on model feature selection,RF showed the most significant enhancement after Bayesian hyperparameter optimization algorithm.The results showed that the Bayesian hyperparameter optimization algorithm was indeed able to improve most of the models’prediction effect.The results of the study provide new methods and ideas for rapid non-destructive testing of tea,which is of great significance for protecting the legitimate rights and interests of consumers and maintaining order in the tea market.