Quantitative analysis of adulterated pear juice based on attenuated total reflectance mid-infrared spectroscopy and bagging PLS

Abstract

The methods of pear juice adulteration have evolved from simple adding water to blending based on characteristic spectrum of juice, which greatly increases the difficulty of identification. Herein, a rapid and convenient analytical method based on attenuated total reflectance (ATR) mid-infrared (MIR) spectroscopy and bagging partial least square (PLS) was developed. A total of 63 samples consisted of three kind of adulteration were designed to simulate actual counterfeiting of pear juice. All samples are roughly evenly divided into the calibration and test sets. Three kinds of PLS models including full-spectrum PLS, local PLS with pre feature selection by the ReliefF algorithm and bagging PLS were trained and compared. On the independent test set, the bagging PLS model achieved the best performance, followed by the local PLS model. Specifically, the bagging PLS model with 40 members achieved a root-mean squared error of prediction (RMSEP) of 3.24 %, a 20 % decrease compared to the full spectrum PLS model. Moreover, the error of bagging PLS model on the calibration and test sets is closer, and there is no over-fitting phenomenon, indicating that its robustness is better, with good fitting and generalization. In today's world where computing resources are not a problem, using simple model ensemble is highly desirable.