Fourier-Transform Mid-Infrared Spectroscopy Coupled with Chemometrics for the Characterization of Commercial Coffee Categories

In Brazil, roasted coffee is marketed under ABIC quality categories (Traditional, Superior, and Gourmet), which do not specify the proportions of Arabica and Conilon in blends. This study investigated the relationship between coffee species and commercial categories and developed a chemometric approach to quantify Arabica-Conilon blends using Fourier-transform mid-infrared (FT-MIR) spectroscopy.

Twenty-two commercial coffee samples from the Gourmet, Superior, and Traditional categories were analyzed, including products labeled as 100% Arabica, 100% Conilon, or blends, alongside uncategorized capsule coffees. Samples were analyzed by attenuated total reflectance FT-MIR (ATR-FT-MIR) using a Bruker Alpha-P spectrometer (4000-400 cm⁻¹, 32 scans). Four preprocessing workflows combining baseline correction, standard normal variate (SNV), Savitzky-Golay first derivatives, and generalized least squares weighting (GLSW) were evaluated across selected spectral regions (3000-2800, 1800-1500, and 1300-900 cm⁻¹).

Principal component analysis (PCA), partial least squares discriminant analysis (PLS-DA) with nested leave-one-group-out cross-validation, and parallel partial least squares regression (PLSR) and extreme gradient boosting (XGBoost) models were employed. Variables Importance in Projection (VIP) scores were used to identify spectral regions contributing to classification and regression.

Label information revealed an association between commercial category and declared botanical composition: all Gourmet products were labelled as 100% Arabica, Superior samples were predominantly Arabica, and Traditional products included pure Conilon coffees or blends with unspecified proportions. The best-performing PLS-DA workflow, combining asymmetric least-squares baseline correction, SNV, Savitzky-Golay derivatives, and GLSW, achieved an external-test accuracy of 0.720, macro-F1 score of 0.725, and Matthew’s correlation coefficient of 0.588, indicating measurable but incomplete discrimination among commercial categories. For independent physical mixtures, PLSR outperformed XGBoost, achieving an R² of 0.972 and an RMSE of 5.94 percentage points. VIP analysis highlighted spectral regions associated with lipids (approximately 1740 and 2920 cm⁻¹), caffeine (approximately 1240, 1650, and 2960 cm⁻¹), and chlorogenic acids as important contributors to model predictions. Capsule coffees occupied a distinct region in the PCA score plot relative to Gourmet Arabica samples, suggesting additional spectral variability associated with product formulation, processing, or storage.

Overall, ATR-FT-MIR combined with chemometric modelling can support the characterization of commercial roasted coffees and the estimation of Arabica-Conilon proportions under the conditions evaluated. However, commercial category discrimination was incomplete, and spectral differences were not exclusively attributable to botanical composition. Further validation using larger and more diverse calibration and independent test sets, encompassing different cultivars, origins, roasting conditions, and product formats, is required before the method can be applied reliably to routine quality control or the investigation of undeclared botanical composition.

This work was financed in part by the Coordenação de Aperfeiçoamento de Pessoal de Nível Superior - Brasil (CAPES) - Finance Code 001.