Quantitative determination of protein content in chocolates using near infrared reflectance spectroscopy with GSVR method

Abstract

This article utilizes near infrared reflectance spectroscopy (NIRS) technology to quantificationally analyze protein content of chocolate, using genetic support vector regression (GSVR) to build spectrum calibration model. GSVR first adopts genetic algorithm to select the efficiency wavelength regions, and then applies linear support vector regression (SVR) to establish a calibration model with the chosen wavelength regions. At the same time SVR method is used to make calibration model for whole spectrum as comparison. 160 samples of 8 typical varieties of chocolates are selected in the experiment. 128 samples are used to train and the remainders are predicting samples. 12 GSVR calibration models and a SVR calibration model with the whole spectrum region are established, and all GSVR models' root mean square error in cross validation (RMSECV) and correlative coefficient ( r ) are better than SVR model's. The best GSVR model, whose RMSECV and r are 0.9767, 0.2565, respectively, is picked out among 12 models as final calibration model. In the predicting process, the GSVR model, whose root mean square error (RMSE) and mean error are 0.2454, 0.1968, respectively, is more robust in contrast to SVR whose RMSE and mean error are 0.2933, 0.2165, respectively.