Use of near-infrared spectroscopy and feature selection techniques for predicting the caffeine content and roasting color in roasted coffees.

Near-infrared spectroscopy (NIRS), combined with diverse feature selection techniques and multivariate calibration methods, has been used to develop robust and reliable reduced-spectrum regression models based on a few NIR filter sensors for determining two key parameters for the characterization of roasted coffees, which are extremely relevant from a quality assurance standpoint: roasting color and caffeine content. The application of the stepwise orthogonalization of predictors (an "old" technique recently revisited, known by the acronym SELECT) provided notably improved regression models for the two response variables modeled, with root-mean-square errors of the residuals in external prediction (RMSEP) equal to 3.68 and 1.46% for roasting color and caffeine content of roasted coffee samples, respectively. The improvement achieved by the application of the SELECT-OLS method was particularly remarkable when the very low complexities associated with the final models obtained for predicting both roasting color (only 9 selected wavelengths) and caffeine content (17 significant wavelengths) were taken into account. The simple and reliable calibration models proposed in the present study encourage the possibility of implementing them in online and routine applications to predict quality parameters of unknown coffee samples via their NIR spectra, thanks to the use of a NIR instrument equipped with a proper filter system, which would imply a considerable simplification with regard to the recording and interpretation of the spectra, as well as an important economic saving.