Prediction of Moisture and Caffeine Contents of the Roasted Coffee Beans (Coffea liberica Hiern) using NIRS and PLSR-MLR Models

Abstract

Near-infrared spectroscopy was assessed for the prediction of moisture content (MC) and caffeine content (CC) of ground “Barako” roasted coffee. Individual models were developed using a chemometric analysis of the NIR spectra (900-1700 nm). Partial least squares regression (PLSR) cross-validation and validation results showed that the MC models could be used for at least quality assurance applications. However, the CC model for PLSR cross validation can only be used for rough screening and approximate calibration applications due to low RPD (2.000) and R2 (0.755) values. The results for the validation models of CC obtained lower RPD (0.220) and R2(0.136) that it did not pass for any use or application. The results of the PLSR modeling identified significant wavelengths based on the regression coefficient and variable importance of projection. These wavelengths were used to develop multiple linear regression (MLR) models. MC model, with 3 wavelengths, was suitable for most research applications with an RPD = 2.600 and R2 = 0.851. CC model, with 8 wavelengths, did not pass for any use or application due to poor predictive performance (RPD = 1.378, R2 = 0.471). The results showed that only the MC models can be used for quality assessment of roasted coffee.