Research On Quantitative Prediction Model Of Main Content In Green Tea Processing Based On Hyperspectral Imaging

Tea industry is a pillar industry in many tea related regions in China,and it is also an ecological industry in the rural revitalization strategy in China.As the largest tea producing and consuming country in the world,China’s green tea production and consumption ranks first among the six major categories of tea,accounting for up to 70%.There is a large market for high-quality green tea,but there are no scientific and unified standards for the processing technology of green tea in many regions of China.The determination of green tea content is mostly based on traditional chemical methods.The state and quality of raw materials during the processing of green tea mainly rely on experience and human judgment,and it is not possible to monitor the content of content in real time.Newly developed equipment and methods also have the problem of insufficient accuracy.In order to further improve the processing quality of green tea in response to the above existing problems,this paper prepared green tea samples under four processes of withering,fixation,rolling,and drying,collected biochemical and hyperspectral data of green tea,and conducted research on the optimization of green tea processing parameters and the establishment of quantitative prediction models for the content of green tea during processing,which has important significance for the production of high-quality green tea.The main research contents are as follows:(1)Biochemical data during the processing of green tea was obtained and the processing technology was optimized.Green tea was processed under multi gradient processing parameters(withering time,fixation temperature,rolling pressure,and drying temperature)during the withering,fixation,rolling,and drying processes.Green tea raw materials were taken as experimental samples and the contents of tea polyphenols,amino acids,caffeine,and water extracts were determined.The optimal processing parameters were determined as follows:withering time 3 to 4 hours,fixation temperature 260℃,rolling pressure 220 N,and drying temperature 180℃.(2)Hyperspectral data during the processing of green tea were collected and preprocessed.Based on the constructed hyperspectral image acquisition system,the hyperspectral images of green tea samples were obtained,and after black and white correction and abnormal sample deletion,the green tea hyperspectral data were preprocessed using smoothing algorithms,multivariate scattering correction,standard normal transformation,and differential processing to eliminate interference such as electrical noise,light scattering,and baseline drift.The partial least squares regression model was established for comparison,The preprocessing methods used were determined as follows:smoothing algorithm(SG),multivariate scattering correction(MSC),and second derivative processing(2Der).(3)The characteristic wavelengths of hyperspectral data were extracted and a set of characteristic wavelengths was obtained.In order to improve modeling efficiency and ensure the accuracy and stability of model prediction,it is necessary to perform wavelength dimension reduction on the hyperspectral data pretreated with green tea.Three methods,including continuous projection method(SPA),principal component analysis(PCA),and correlation analysis(CA),were used to extract characteristic wavelengths from the hyperspectral data pretreated with MSC and MSC+2Der,respectively,to obtain multiple sets of characteristic wavelengths for establishing content prediction models(4)The prediction model of green tea content was constructed and optimized.After the green tea sample set was divided into a prediction set and a correction set by the Kennard Stone(KS)method,the partial least squares(PLSR),support vector machine(SVR),least squares support vector machine(LSSVR),and back propagation(BP)neural network prediction models for the four inclusions were established.Then the particle swarm optimization(PSO)algorithm is introduced to optimize the modeling parameters of the optimal model LSSVR to establish a PSO-LSSVR prediction model.The results show that the R~2 of the PSO-LSSVR prediction model is greater than 0.9,the maximum RMSE is only around 0.005,the maximum MAPE is not more than 2%,and the RPD is also greater than 3.Compared to LSSVR,the prediction ability is better and the prediction accuracy is higher.