Variable Selection Applied to the Development of a Robust Method for the Quantification of Coffee Blends Using Mid Infrared Spectroscopy

Abstract

This paper combined attenuated total reflectance Fourier transform infrared spectroscopy (ATR-FTIR), multivariate calibration with partial least squares (PLS), and different variable selection methods for the development of models to determine Robusta-Arabica coffee blends in the analytical range from 0.0 to 33.0% w/w. Ground samples of different origins were roasted at three different levels: light, medium, and dark. Specific models were built for each roasting level, and a robust model was also obtained including all the samples. Mid infrared spectra were recorded in the wavenumber range between 4000 and 800 cm−1 for the 120 samples used in the models. Four variable selection methods were tested: genetic algorithm (GA), ordered predictors selection (OPS), successive projections algorithm (SPA), and interval PLS (iPLS). The best results were obtained using GA and OPS, decreasing root mean square errors of prediction (RMSEP) in 44–68% as compared to full spectra models. The best robust model was obtained with OPS, providing RMSEP of 1.8% w/w. The number of selected variable in the optimized models varied from 6.5 to 17.0% of the total number of original variables. This demonstrated the importance of selecting a limited number of wavenumbers richer in information specifically related to the analytes. All the methods were validated by estimating appropriate figures of merit and considered accurate, linear, sensitive, and unbiased.