**3. Results**

#### *3.1. ATR-FTIR and NIR Analysis*

The spectra presented in Figure 1 represent the average FTIR spectra of four classes of samples grouped according to their score of sensory quality: 81–83; 84–86; 87–89; 90+. The two bands at the 2900–2850 cm<sup>−</sup><sup>1</sup> range are attributed to C-H vibrations of the bonds present in lipid and caffeine molecules [16]. The marked 1750 cm<sup>−</sup><sup>1</sup> and 1650 cm<sup>−</sup><sup>1</sup> regions are attributed to carbonyl (C=O) vibration and C=C bonds, attributed to carbohydrates and lipids, respectively [20].

**Figure 1.** Average FTIR spectra obtained for roasted coffee (colors are related to sensory quality scores).

The bands in the 1650–1600 cm<sup>−</sup><sup>1</sup> range have been previously reported in association with caffeine [4], and were employed in previous studies for quantitative analysis of this substance. The 1680–1630 cm<sup>−</sup><sup>1</sup> range has been found to be associated with vibrations in the carbonyl amide group [4] and also to the presence of trigonelline. The latter is usually decomposed into pyrroles and pyridines during roasting. Pyridines are some of the substances that are responsible for the characteristic aroma of roasted coffee [21].

A significant number of bands can be observed between 1500 and 900 cm<sup>−</sup>1. Carbohydrates have several absorption bands in the region between 1400 and 900 cm<sup>−</sup>1, also called "fingerprint region", because it concentrates several relevant bands. The band at 1146 cm<sup>−</sup>1, has been linked to polysaccharides in previous studies, specifically to the C-O-C stretching of the glycosidic link in the cellulose molecule [4]. Bands in this have also been attributed to amino acids and proteins [22]. Nonetheless, an accurate chemical assignment of bands in this region is still a challenge because of highly coupled vibration modes of polysaccharide backbones [4]. The bands at the 1450–1150 cm<sup>−</sup><sup>1</sup> range have been reported in association with the presence of chlorogenic acids [16,22]. The band at 930 cm<sup>−</sup><sup>1</sup> has been previously reported in association with residues of 3,6-anhydro-galactopyranose [23] resulting from the thermal degradation of polysaccharides, such as galactomannan and arabinogalactan. The levels of chlorogenic acid and trigonelline as well as carbohydrate content will change significantly with roasting, so variations in the fingerprint region of the spectra are expected [14,24].

Figure 2 shows the average NIR spectra of the samples, with different colors being associated with the sensory quality score. The most relevant bands present in the data are as follows: 1100–1250 nm (associated to CH, C-H2, and CH3 overtones from proteins,

lipids, caffeine, and organic acids), and 1300–1490 nm (first overtones of RN-H of proteins, first overtones of OH of water and acids) [25]. The band in the 1900 nm region is associated with the combination of O-H angular stretching and deformation, related to the presence of water [26]. The region of 1208–1236 nm is the second bond overtone of C-H, C-H2, and C-H3, as well as the 1700–1720 nm region, which is related to the first overtone of the same carbon and hydrogen bonds, and C-H bonds linked to aromatic rings [26].

**Figure 2.** Full NIR spectra (1000–2000 nm) obtained for roasted coffee (colors are related to sensory quality scores.

#### *3.2. Partial Least Squares Regression (PLS)*

Figures 3 and 4 show the plots of measured vs. estimated values obtained for the models based on the spectra (estimated) in comparison to the quality scores provided by the Q-graders (measured). The models' parameters are shown in Table 1. The FTIR model was built with two latent variables that were able to explain 99.71% and 81.2% of the accumulated variance in the spectra and sensory data, respectively. Both the values obtained for RMSEP and RMSEC were 0.23%, whereas calibration and validation coefficients were 0.99 and 0.97, respectively. In comparison, the NIR model used three latent variables that explained 90.4% of spectra data variance and 54.05% of the score (sensory) data. The RMSEC value was 0.50% and RMSEP value of 0.52%, and both the calibration (Rc) and validation (Rv) correlation coefficients were 0.98. Although both NIR and FTIR were able to provide good predictions, the FTIR results were slightly more accurate, given the smaller values for RMSEC and RMSEP and slightly higher values for calibration and validation in comparison to NIR. The potential of FTIR as a tool to classify specialty coffees was reported by Belchior et al. [3]. The comparison of both techniques highlights the efficiency of NIR as well. Nonetheless, despite its grea<sup>t</sup> potential in food analysis, the interpretation of the spectra in NIR analysis is challenging due to its broadband nature, which consists of overlapping overtone and combination bands [7,9]. The use of NIR-based models represents a new approach in comparing the chemical data with the SCA protocol for coffee classification. Although some studies have reported the use of NIR to evaluate coffees [9,25,27], the discrimination between high quality samples as well as the comparison with the SCA classification shown in this study confirms the potential of this method,

providing the coffee industry with a good perspective for using different spectroscopy tools to evaluate coffee quality. Although previous studies [26,27] were able to show that NIR can be used to predict specific coffee sensory parameters (body, acidity, flavor, aftertaste, etc.), this is the first study that extends this application to quantifiable quality scores using an internationally accepted sensory protocol, thus confirming the potential of this method for coffee quality evaluation.

**Figure 3.** Experimental (black circles) vs. predicted values (pink triangles) obtained by the models based on FTIR spectra.

**Figure 4.** Experimental (black circles) vs. predicted values (pink triangles) obtained by the models based on NIR data.


**Table 1.** Comparison of the PLS models for both FTIR and NIR techniques.

RMSEC = root mean square error of calibration; RMSEP = root mean square error of validation; Rc = calibration correlation coefficient; Rv = validation correlation coefficient.

Figures 5 and 6 show the Variable Importance of Projection (VIP) scores of the models. A VIP score is a measure of a variable's importance in the PLS model and is calculated as the weighted sum of squares of the PLS weights, taking into account the amount of explained variance in each extracted latent variable (dimension). Thus, VIP scores above 1 are a typical rule for selecting relevant variables in a given model. An evaluation of the FTIR VIP scores (Figure 5) indicates that the entire spectrum affected the coffee classification in association with the SCA classification. The fact that bands in the whole wavenumber range were relevant in terms of sample classification indicates that many substances that are present in the coffee beverage have significant impact on the sensory profile. Besides coffee being a complex food matrix, several variables that affect the coffee processing chain processes (cultivation, harvesting, post-harvesting, storage, roasting, grinding, and extraction) will impact the final product and affect sensory variations that can be perceived by the Q-graders. Although roasting conditions are consistent and established in terms of the SCA protocol, there still can be variations in the roasting profile (environmental conditions, type of equipment, etc.) [28,29].

**Figure 5.** VIP Scores of the PLS models based on FTIR data.

**Figure 6.** VIP Scores of the PLS models based on NIR data.

VIP scores obtained for the NIR model (Figure 6) show the bands 1176, 1749, and 1950 nm as the most relevant in predicting the coffee scores, being related to the second overtones of CH, C-H2, and CH3, first overtones of CH and C-H2, first overtones of OH, RCO2R, and CONH2, and second overtones of C=O [26]. These regions were assigned by Ribeiro et al. [26] as related to the sensory characteristics of the attributes: taste, cleanliness, and body of the beverage. The region comprised between 1156-1172 nm is attributed to caffeine, and 1738–1755 nm to the presence of lipids in the samples. The region between 1937–1959 nm are related to the ACG and water content of the samples. The beverage cleanliness, regarding the quality of body, is a relevant attribute evaluated in coffee and responsible for higher scores, reinforcing the feasibility of NIR data in adding more confidence to the results.

Schenker and Rothgeb [30] stated that the roasting process can be divided into three stages: drying, Maillard reactions, and development. Therefore, the sensory profile of the roasting coffee will be directly affected by roasting time because it is directly related to the specific phase of the roasting process. This will affect the final coffee composition in terms of several components and reactions, including chlorogenic acids and their derivatives, sugar caramelization, organic acids, volatiles, lipid migration, and melanoidin production; such composition will have a direct effect on the sensory profile [18,31,32]. Therefore, the possibility of validating the sensory analysis performed by Q-graders by using spectroscopic methods is quite relevant. The results obtained in this study are promising for the classification of specialty coffees and confirm the potential of both NIR and FTIR as fast and efficient alternatives for the task at hand. Furthermore, the results are of high interest to the coffee industry, bringing more confidence to the trading routine, given possible inconsistencies between classification of the same samples by sellers and buyers.

## **4. Conclusions**

Spectroscopy-based methods, FTIR and NIR, were shown to be appropriate tools for confirming and predicting score classifications given by Q-graders to roasted specialty coffee samples. The results are promising from the chemometrics standpoint, with models presenting high values for calibration and validation correlation coefficients for both techniques, showing that NIR is also a good tool for predicting coffee quality. It is noteworthy that, even with all samples being of high quality, it was possible to discriminate the nuances in sensory profile. Although the analysis of the whole FTIR spectra of coffee seems to be slightly more efficient from a scientific point of view, NIR spectra also provided

robust results related to relevant chemical parameters that define specialty coffee, such as the cleanliness of the beverage. NIR seems promising for routine analysis of specialty coffees, given its simplicity and the possibility of using portable equipment. Therefore, both techniques can be used to confirm and verify the coffee quality scores associated with the Q-graders assessment. As a result, the coffee industry would increase confidence in trading purposes, producing more consistent results. Nonetheless, further studies are needed in order to increase model robustness and applicability, given the intrinsic variations in coffee samples associated to geographical origin, edaphoclimatic conditions, cultivation, and processing techniques as well as variations in roasting parameters. The variability in roasting conditions and equipment in the case of commercially available roasted coffee samples, and the fact that the present methodology was not validated for such conditions, is noteworthy. One of the difficulties in using sensory analysis in the case of the coffee beverage is the need for strict control of roasting conditions in order to guarantee that the tasters will be able to perceive the flavors appropriately. The SCA protocol and the models herein used will be able to correctly classify specialty coffees prior to roasting, but are not suitable for samples that are already acquired as roasted coffees with varying degrees of roast.

**Supplementary Materials:** The following supporting information can be downloaded at: https: //www.mdpi.com/article/10.3390/foods11111655/s1, Table S1: Coffee sample provenance; sensory scores and description provided by the Q-graders.

**Author Contributions:** Conceptualization, V.B. and A.S.F.; Data curation, V.B. and B.G.B.; Formal analysis, V.B., B.G.B. and A.S.F.; Funding acquisition, A.S.F.; Methodology, V.B., B.G.B. and A.S.F.; Project administration, A.S.F.; Supervision, A.S.F.; Writing—original draft, V.B. and B.G.B.; Writing— review and editing, V.B. and A.S.F. All authors have read and agreed to the published version of the manuscript.

**Funding:** A.S.F. acknowledges financial support from the Brazilian National Council for Scientific and Technological Development, CNPq (Grant # 310456/2021-5). V.B. acknowledges her scholarship from Coordenação de Aperfeiçoamento de Pessoal de Nível Superior, CAPES.

**Institutional Review Board Statement:** The study was conducted in accordance with the Declaration of Helsinki, and approved by the Brazilian National Ethics Committee (CAAE code 56961316.0.0000.5093 aproved in 8 August 2016).

**Informed Consent Statement:** Informed consent was obtained from all subjects involved in the study.

**Data Availability Statement:** Data is contained within the article (or supplementary material).

**Acknowledgments:** IKAWA® Sample Roaster Pro (London, UK) supplied by Macchine Per Caffè Ltda (São Paulo, São Paulo, Brazil).

**Conflicts of Interest:** The authors declare no conflict of interest.
