Evaluation of Least Squares Support Vector Machine Regression and other Multivariate Calibrations in Determination of Internal Attributes of Tea Beverages

Abstract

This research aimed to explore the relationship between internal attributes (pH and soluble solids content) of tea beverages and diffuse reflectance spectra. Three multivariate calibrations including least squares support vector machine regression (LSSVR), partial least squares (PLS), and radial basis function (RBF) neural network were adopted for development of internal attributes determination models. Ten kinds of tea beverages including green tea and black tea were selected for visible and near infrared reflectance (Vis/NIR) spectroscopy measurement from 325 to 1,075 nm. As regard the kernel function, least squares–support vector machine regression models were built with both linear and RBF kernel functions. Grid research and tenfold cross-validation procedures were adopted for optimization of LSSVR parameters. The generalization ability of LSSVR models were evaluated by adjusting the number of samples in the training set and testing set, and sensitive wavelengths that were closely correlated with the internal attributes were explored by analyzing the regression coefficients from linear LSSVR model. Excellent LSSVR models were built with r = 0.998, standard error of prediction (SEP) = 0.111, for pH and r = 0.997, SEP = 0.256, for soluble solids content, and it can be found that the LSSVR models outperformed the PLS and RBF neural network models with higher accuracy and lower error. Six individual sensitive wavelengths for pH were obtained, and the corresponding pH determination model was developed with r = 0.994, SEP = 0.173, based on these six wavelengths. The soluble solids content determination model was also developed with r = 0.977, SEP = 0.173, based on seven individual sensitive wavelengths. The above results proved that Vis/NIR spectroscopy could be used to measure the pH and soluble solids content in tea beverages nondestructively, and LSSVR was an effective arithmetic for multivariate calibration regression and sensitive wavelengths selection.