Classification and quantification of defective and non-defective coffees by FTIR and NIR spectroscopy

Abstract

A major parameter directly related to coffee quality is the presence of defective beans, which impart negative sensory aspects to the beverage. The defects that contribute the most to the depreciation of the beverage quality are black, sour and immature beans. The conventional method used to assess the quality of roasted coffees is based on sensory evaluation, which, although reliable, is time-consuming and requires trained cupper experts. In view of the aforementioned, the objective of the present study was to evaluate the potential of FTIR and NIR spectroscopy as practical techniques to assess the quality of coffees based on the presence of defective beans. Coffee beans were manually sorted into five classes: black, dark sour, immature, light sour and non-defective. Each of the coffee classes was roasted at three temperatures (220 °C, 235 °C and 250 °C) and to three roasting degrees (light, medium and dark) obtaining nine roasting conditions. Roasted coffee samples were ground, sieved and analyzed by DRIFTS, ATR-FTIR and NIR for a classification study. Results from PCA indicated that based on DRIFTS spectra, coffee samples could be discriminated into four major groups: (a) non-defective, (b) black, (c) dark sour and (d) light sour, with immature beans scattered among the sour samples. ATRFTIR provided the discrimination of the coffee samples, although not clearly, into two groups: (a) non-defective and light sour and (b) black, dark sour and immature, and NIR provided the discrimination into three major groups: (a) non-defective, light sour and immature, (b) dark sour, and (c) black. At all cases the variance among the samples led to the discrimination of the coffees primarily by their classes, regardless of roasting degree. Classification models for DRIFTS spectra were developed by LDA while classification models for ATR-FTIR and NIR were developed by Elastic net. High percentages of correct classification, up to 100%, were achieved with each of the techniques employed. The discriminating variables that contributed to the correct classification of the samples from the Elastic net models, for ATR-FTIR and NIR data, were extracted and provided the following interpretation of the models: (a) nondefective coffee was directly related to high levels of carbohydrates and lipids and lower levels of proteins and/or amino acids and caffeine; (b) light sour coffee was related to high levels of carbohydrates and caffeine; (c) dark sour coffee was directly associated with high levels of aliphatic acids and low levels of lipids; (d) black coffee was related to high levels of proteins and/or amino acids and low levels of lipids; and (e) immature coffee was related to high levels of proteins and/or amino acids and caffeine and low levels of lipids. In a second part of this study, blends of defective in admixture with non-defective coffee, with %defects ranging from 0% to 30% in steps of 3%, were produced and analyzed by ATR-FTIR and NIR for a quantification assay. PLSR was used to construct the models that provided satisfactory results. RMSEP values as low as 2.6% and R2 values as high as 0.956 in the validation set were achieved. Overall, NIR overcame ATR-FTIR in terms of robustness and accuracy.