Prediction of a wide range of compounds concentration in raw coffee beans using NIRS, PLS and variable selection

Abstract Near infrared spectroscopy coupled to chemometrics has been proposed as a low cost, rapid and eco-friendly methodology in both off-line and on-line analyses of coffee beans and coffee beverages. However, there are some methodological restrictions regarding its use to quantify chemical constituents in raw coffee beans. In this sense, attention can be drawn to the slight variability of the reference data needed for the construction of multivariate models and the number of analyses needed for the reference method. To overcome these limitations and favor the spectroscopy use, innovations were introduced in the methodological approach for quantifying caffeine, trigonelline and 5-caffeoylquinic acid (5-CQA). Novel mixtures and doping of matrices of different species as well as variable selection processes were proposed. Partial least squares regression (PLSR) was used as a multivariate analysis to build the calibration model for each compound. Using 7, 6 and 5 latent variables, the prediction models constructed for caffeine, trigonelline and 5-CQA contents resulted in a RMSEP of 0.08, 0.07 and 0.27 and rvc of 0.98, 0.96 and 0.96, respectively. In addition, a total of 46 wavelength regions were selected and discussed as important markers for predicting the compounds concentration.