Characterization of Mexican coffee according to mineral contents by means of multilayer perceptrons artificial neural networks

Characterization of Mexican coffee according to mineral contents by means of multilayer perceptrons artificial neural networks

Journal of Food Composition and Analysis 34 (2014) 7–11 Contents lists available at ScienceDirect Journal of Food Composition and Analysis journal h...

656KB Sizes 0 Downloads 57 Views

Journal of Food Composition and Analysis 34 (2014) 7–11

Contents lists available at ScienceDirect

Journal of Food Composition and Analysis journal homepage: www.elsevier.com/locate/jfca

Original Research Article

Characterization of Mexican coffee according to mineral contents by means of multilayer perceptrons artificial neural networks ˜ iz-Valencia a,*, Jose´ M. Jurado b, Silvia G. Ceballos-Magan ˜ a c, A´ngela Alca´zar b, Roberto Mun a Julio Herna´ndez-Dı´az a b c

Facultad de Ciencias Quı´micas, Universidad de Colima, Carretera Colima-Coquimatla´n km 9, 28400, Coquimatla´n, Colima, Mexico Department of Analytical Chemistry, Faculty of Chemistry, University of Seville, c/ Profesor Garcı´a Gonza´lez 1, 41012 Seville, Spain Facultad de Ciencias, Universidad de Colima, c/ Bernal Dı´az del Castillo 340, 28045 Colima, Mexico

A R T I C L E I N F O

A B S T R A C T

Article history: Received 3 December 2012 Received in revised form 18 April 2013 Accepted 3 February 2014

The content of Ca, Cu, Fe, K, Mg, Mn, Na and Zn has been determined in Mexican roasted coffee beans from four producing states by means of inductively coupled plasma optical emission spectrometry (ICPOES). The concentrations of these elements were used to differentiate the coffee growing area. Kruskal– Wallis test highlighted significant differences between metals in samples from the four origins. Principal component analysis was used to visualize the natural trends of data distribution for the considered groups. Forward stepwise linear discriminant analysis (LDA) was used to differentiate coffee origins as well as to find out the best chemical descriptors (Ca, K, Mn, Mg, Na and Zn). The overall sensitivity and specificity of LDA were 81% and 94%, respectively. These results were improved when a multilayer perceptron artificial neural networks model was applied, allowing the differentiation of Mexican roasted coffees with 93% prediction ability and 98% specificity. ß 2014 Elsevier Inc. All rights reserved.

Keywords: Food analysis Food composition Metal content in coffee Geographical authentication Pattern recognition Coffee Linear discriminant analysis Artificial neural networks

1. Introduction World coffee consumption has seen strong growth over the last ten years, reaching an estimated 137.9 million bags in 2011 (ICO-a, 2012). Consequently, geographical characterization of coffee is an interesting topic due to the economical importance for producing countries, the establishment of products with certified origins or agricultural practices, and the preference of consumers for certain products with differentiating characteristics (Barham and Weber, 2012). There are two commercially used coffee species, namely Coffea arabica L. and Coffea canephora, known as arabica and robusta, respectively (Carrera et al., 1998). Mexico is the worlds’ 5th largest arabica coffee producer, with Chiapas, Colima, Oaxaca and Veracruz as the main growing states (Historia del cafe, 2012). The arabica coffees beans command higher prices (by >200%) than the respective robusta coffee beans. Nevertheless, the price also depends on the geographic origin (ICO-b, 2012). Taking into account this fact, several works have focused on the discrimination

* Corresponding author. Tel.: +52 3123161163; fax: +52 3123161163. ˜ iz-Valencia). E-mail address: [email protected] (R. Mun http://dx.doi.org/10.1016/j.jfca.2014.02.003 0889-1575/ß 2014 Elsevier Inc. All rights reserved.

of coffee species. The composition also depends on environmental factors and it is becoming more important to differentiate among coffee origins to assess coffee quality (Choi et al., 2010). Chemical composition can be used to differentiate between coffee varieties and geographical origin by means of pattern recognition techniques. Most of the literature on this topic has been carried out with the aim of differentiating between arabica and robusta species, as the literature about coffee origin discrimination is very scarce. Major growing regions have been differentiated using the contents of mineral elements (Anderson and Smith, 2002; Akamine et al., 2010), phenolic and methylxanthine (Alonso-Salces et al., 2009), metabolomic profiles (Choi et al., 2010), volatile composition (Risticevic et al., 2008), fatty and chlorogenic acids (Bertrand et al., 2008) and elemental profiles and stable isotopes (Santato et al., 2012). The purpose of this study was to use inductively coupled plasma optical emission spectrometry (ICP-OES) multi-element analysis combined with chemometrics for differentiating the geographical origin of Mexican coffee beans from the states of Chiapas, Colima, Oaxaca and Veracruz (Fig. 1). The presented work gives important identity information about the coffee produced in Mexico contributing to the development of the Mexican agribusiness. In addition, there is no reported methodology for differentiating Mexican coffee

R. Mun˜iz-Valencia et al. / Journal of Food Composition and Analysis 34 (2014) 7–11

8

Table 1 ICP-OES operating conditions.

Fig. 1. Studied Mexican coffee production areas.

beans. The proposed method consists of: (i) the mineralization of coffee samples using a microwave digestion system; (ii) the determination of, Ca, Cu, Fe, K, Mg, Mn, Na and Zn by means of ICP-OES; (iii) the application of Pattern recognition techniques, such as linear discriminant analysis (LDA) and multilayer perceptron artificial neural networks (MLP-ANN) in order to differentiate their geographical origin. 2. Materials and methods 2.1. Chemicals and reagents Water (HPLC grade) from Fermont (Nuevo Leon, Mexico), trace metal grade 70% nitric acid from J.T. Baker (Estado de Mexico, Mexico), 30% hydrogen peroxide and 1000 mg/L elemental stock standard solutions of K, Ca, Fe, Na, Mg, Mn, Cu and Zn purchased from Sigma–Aldrich (St. Louis, MO, USA) were used. A single or a mixture of these elements was prepared daily by diluting the stock solution with ultrapure water in 5% (v:v) nitric acid. Prior to use, all glassware and polypropylene flasks were washed with 10% (v:v) nitric acid and rinsed with ultrapure water. 3. Coffee samples All samples used for this study were roasted coffee beans belonging to C. arabica L. variety. A total of 51 commercial samples of coffee from states of Mexico (Fig. 1): Chiapas (n = 12), Colima (n = 24), Oaxaca (n = 6) and Veracruz (n = 9) were purchased from local producers. Samples were 100% arabica and their geographical authenticity was established according to the information given by the local suppliers. 3.1. Samples analysis Prior to the determination of Ca, Cu, Fe, Na, K, Mg, Mn and Zn, coffee samples were dried at 103 8C for determination of their moisture (International Organization of Standardization, 1994). An Optima 7000 ICP-OES spectrometer with dual view configuration (Perkin Elmer, Waltham, Ma, USA) and the WinLab32 software package was used for determining metal content. The operational parameters are shown in Table 1. The mineralization of coffee samples was accomplished using a QLab 6000 closed vessel microwave digestion system (Questron Technologies, Mississauga, ON, Canada) equipped with 200 mL Teflon PFA vessels. The used protocol was adapted from that proposed by Fernandes et al. (2005). Briefly, a 250 mg milled sample (from an agate mortar) was weighed in a PTFE digestion vessel and mixed with 1 mL of 30% H2O2 and 3 mL

Parameter

Condition

Power Plasma flow Auxiliary flow Nebulizer flow Replicates Spray chamber Nebulizer Emission lines (nm)

1450 kW 17 L/min 0.20 L/min 0.55 L/min 3 Cyclonic Meinhard K 766.490, Mg 285.213, Ca 317.933, Fe 238.204, Na 589.592, Mn 257.610, Cu 327.393, Zn 206.200

of 70% HNO3. Subsequently, the vessel was placed on the microwave turntable to digest the coffee sample. The microwave oven temperature program consists of increasing from ambient temperature to 160 8C in 10 min, holding at 160 8C for 10 min before cooling to ambient temperature. The resulting solution was diluted to 25 mL with ultrapure water before the ICP-OES determination in triplicate. 3.2. Data analysis A data matrix with eight columns (determined elements) and fifty one rows (analyzed samples) was built for chemometric calculations. Kruskal–Wallis test was used to highlight potential discriminant variables according to the statistical differences found between coffee origins. Principal component analysis (PCA) was used to previsualize data trends. Linear and non-linear pattern recognition techniques, such as LDA and MLP-ANN, respectively, were applied to carry out the classification of coffee samples according to their geographical origin. Data processing was made by using the statistical package Statistica 8.0 (StafsoftTM, Tulsa, OK, USA). Auto scaled data were used in all calculations. 4. Results and discussion 4.1. Method performance The performance characteristics of the method, such as trueness, linearity on the calibration range, limit of detection (LOD) and quantification (LOQ), and precision (repeatability and intermediate precision) were assessed. The trueness of the method was evaluated by means of recovery assays. A control sample was fortified in n levels (n = 3) corresponding to a low, medium and high level of concentration, covering the 80%, 100% and 120% of the mean expected value (Gonza´lez et al., 2005). Each level was analyzed in triplicate. The global recoveries (R) for each element with their expanded uncertainties (U) were calculated (Jurado et al., 2007) and the results are shown in Table 2. As can be seen, recoveries do not Table 2 Recovery, linearity (% L), sensitivity (LOD and LOQ), repeatability and intermediate precision and obtained for the analyzed elements. Element

R (%)

%L

LOD (mg/kg)

LOQ (mg/kg)

RSDrepeat (%)

RSDIP (%)

Ca Cu Fe K Mg Mn Na Zn

100  2 99  3 99  2 98  3 100  6 101  3 104  4 101  7

99.5 98.7 99.2 99.0 98.8 99.7 99.8 98.9

0.2 0.2 0.1 0.3 0.4 0.2 0.5 0.1

0.8 0.7 0.4 0.9 1.5 0.3 1.7 0.2

2.0 1.5 1.6 2.4 1.0 2.1 1.0 0.8

7.5 6.2 4.8 5.2 3.4 6.8 3.5 4.4

R (%), mean recovery  expanded uncertainty; % L, linearity; LOD, limit of detection; LOQ, limit of quantification; RSDrepeat, repeatability; RSDIP, intermediate precision; Each sample was analyzed in triplicate.

R. Mun˜iz-Valencia et al. / Journal of Food Composition and Analysis 34 (2014) 7–11

9

Table 3 Medians, minimum and maximum values (in brackets) of metal content in Mexican coffees. Concentrations expressed in mg/kg (dry wt). State

Chiapas

Colima

Oaxaca

Veracruz

All groups

Ca Cu Fe K Mg Mn Na Zn

6135 (833–14,391) 15 (n.d.–59) 95 (n.d.–661) 13013 (12,107–13,753) 1539 (1260–1839) 20 (9–27) 680 (103–1167) 60 (0.9–140)

1018 (707–1495) 13 (n.d.–165) 24 (n.d.–222) 13204 (10,566–15,273) 1559 (1156–1808) 6 (4–8) 427 (115–1265) 30 (10–140)

881 (719–1076) n.d. (n.d.–16) 12 (5–55) 12,814 (12,080–13,595) 1481 (1360–1610) 10 (8–13) 170 (11–330) 5 (0.4–10)

832 (654–1283) 14 (n.d.–84) 33 (11–189) 12,602 (10,735–13,451) 1329 (1286–1608) 13 (9–15) 327 (131–1177) 26 (8–135)

1024 (654–14,391) 13 (n.d.–165) 24 (n.d.–661) 12,911 (10,566–15,273) 1531 (1156–1839) 9 (4–27) 331 (11–1265) 26 (0.4–140)

n.d. not detected; Each sample was analyzed in triplicate.

Table 4 Kruskal–Wallis test results. Element

Ha

Ca Cu Fe K Mg Mn Na Zn

13.52 4.01 4.08 6.88 4.19 40.28 7.99 10.60

Significant differences Chiapas–Colima

a

Chiapas–Oaxaca

Chiapas–Veracruz

X

X

X

Colima–Oaxaca

Colima–Veracruz

X X X

X

Oaxaca–Veracruz

Significant difference for H > 7.82.

significantly differ from 100% as the inequality |R 100|  U holds true for all of the studied elements and accordingly the trueness of the method is assured. Calibration range for Na, K, Ca, and Mg, was in the range of 0– 50 mg/L, whereas for Fe, Cu, and Zn the range was between 0 and 25 mg/L range, and for Mn calibration was in the range of 0– 2.5 mg/L. Linearity in the calibration range was calculated as % L = 100  (1 sb/b), with b being the slope of the nitric acid matched calibration curve and sb its standard deviation (Cuadros et al., 1996). As can be seen in Table 2, linearity of elements presented values higher than 98.7%. The LOD and LOQ were calculated as the concentration corresponding to a signal 3 and 10 times the standard deviation of the blank, respectively. The obtained results are shown in Table 2. The LOD for Na was 0.5 mg/kg and for the remaining elements were in the 0.1–0.4 mg/kg range. The LOQ for Mg and Na were 1.5 and 1.7 mg/kg, respectively. The remaining elements presented LOQ in the range 0.2–0.9 mg/kg. Repeatability was evaluated by the mineralization and analysis of the control sample over a short period of time and without changing any operational condition. The relative standard deviation of repeatability (RSDrepeat) was calculated. On the other hand, intermediate precision (IP) was obtained by analyzing the sample over a long period of time and RSDIP was computed. As can be seen in Table 2, RSDrepeat and RSDIP presented values in the range of 0.8– 2.4% and 3.4–7.5%, respectively. These results are in accordance with those obtained from the Horwitz function (Horwitz, 1982) depending on the analyte concentration level. 4.2. Sample analysis Mexican coffees were analyzed to determinate the contents of Ca, Cu, Fe, K, Mg, Mn, Na and Zn. The ranges and median values, expressed as mg/kg in dry weight (wt) basis, are shown in Table 3. Major compounds are K, Mg, Ca and Na with medians of 12,991 mg/kg, 1531 mg/kg, 1024 mg/kg and 331 mg/kg, respectively. On the other hand, Mn, Cu, Fe and Zn, with median values of 9 mg/kg, 13 mg/kg, 24 mg/kg and 26 mg/kg, respectively, are considered as minor elements. Coffees from the state of Chiapas

present the highest median content for Ca (6135 mg/kg), Na (680 mg/kg), Fe (95 mg/kg), Zn (60 mg/kg) and Mn (20 mg/kg). Coffees from Colima, with median value of 6 mg/kg, exhibit the lowest content of Mn. Oaxaca coffees present the lowest medians for Cu (non detected, n.d.), Fe (12 mg/kg), Na (170 mg/kg) and Zn (5 mg/kg). Veracruz coffees in comparison with coffees from the other states present the lowest medians for K (12,602 mg/kg), Mg (1329 mg/kg) and Ca (832 mg/kg). The Cu, Na, and Zn contents in Veracruz coffees are similar to those found in the Colima coffees. K and Mg contents appear very similar for the four considered origins. Kruskal–Wallis test was applied in order to highlight significant differences in the metal content among coffees from the four Mexican states. An H-value is computed and compared with the tabulated chi-square for a = 0.05 and n 1 degrees of freedom (n = 4). Table 4 shows the computed H-values for each of the eight elements as well as the results of the posthoc analysis. The highest H-value (40.28) was obtained for the content of Mn, showing statistical differences between samples from Colima and the other three considered states. Ca presents an H-value of 13.52, with significant differences between the pairs Chiapas–Oaxaca and Chiapas–Veracruz Zn (10.60) and Na (7.99) present significant differences in the comparison Colima–Oaxaca. Cu, Fe, K, and Mg present non-significant H-values. PCA was performed to visualize data trends to appear separately. PCA obtains new variables called principal components (PCs) as a linear combination of the original variables. The PCs are calculated sequentially and each new PC account as much as possible for the residual variance of the data. A graphical representation of the cases in the space of the first two or three PCs allows the visualization of data trends with a lesser dimensionality (Jolliffe, 2002). In this case, the first three PCs (those with Eigen values greater than 1) were obtained, explaining 54.1%, 18.0% and 13.0% of total variance, respectively. The most contributing variables to PC1 were Ca, Fe, Mg, Na, and Zn. PC2 and PC3 are highly influenced by Mn and K, respectively. Fig. 2 shows the distribution of samples in the space of the three computed PCs. No clear separation is achieved although some trends can be observed. For instance, coffees from Chiapas appear at negative values of PC1, most samples from Colima have positive PC2 values and samples of Veracruz are found at negative values of PC3.

10

R. Mun˜iz-Valencia et al. / Journal of Food Composition and Analysis 34 (2014) 7–11

Fig. 3. Distribution of coffee samples in the space of the three computed DFs. Fig. 2. Distribution of coffee samples in the space of the three first PCs computed from the metal contents.

Coffees from Oaxaca appear in the ranges 2–3, 0–1 and 1 to 0 of PC1, PC2 and PC3, respectively. These trends to differentiation are in accordance to the results of the previous Kruskal–Wallis multiple comparison test. As can be seen, the four variables showing statistical differences, such as Mn, Ca, Zn and Na, also appear as most contributing to the two first PCs. Then, it can be thought that these elements are potentially good chemical descriptors to be used in classification models. As natural separation of sample groups was not observed by PCA, supervised pattern recognition techniques, such as LDA and ANN, were used to solve the classification problem. LDA linearly combines the original variables to obtain discriminant functions (DFs) which separates categories by means of the minimization of the within-class and between-class ratio of the sum of squares (Massart, 1998). Forward stepwise LDA was carried out to compute three DFs selecting the most discriminant variables from the original ones. In stepwise LDA, the model of discrimination is built step by step. At each step all variables are reviewed and evaluated to determine which one contributes most to the discrimination between groups, by increasing the ratio of between-groups and within-group variances. That variable is then included in the model and the process starts again. In this case the obtained model was based on the contents of Ca, K, Mg, Mn, Na, and Zn. Four of these elements, such as Ca, Mn, Na and Zn are the same highlighted by Kruskal–Wallis test and PCA. Mg highly contributes to PC1, which showed differences in the case of Chiapas and Colima. K contributes to PC3, which seems to separate samples from Veracruz and Colima. Then, the potential goodness of these six elements as chemical descriptors is now corroborated by stepwise LDA, and they were subsequently selected for building pattern recognition models. Fig. 3 shows the sample distribution in the space of the three computed DFs. Samples from Colima appears separated from the other classes at negative values of DF1 and positive values of DF2. Coffees from Oaxaca appear in a narrow window, differentiated in the plane of DF1 with values in the 0–2 range and low negative scores in the plane of DF3. Samples from Chiapas and Veracruz overlapped appearing at values bigger than 3 of DF1, positive scores of DF2 and negative of DF3. To validate the model samples were divided into a training set and a test set. The training set, containing 3/4 of cases, was used to build the LDA model, which was applied to the test set, which consist of the remaining 1/4 of the cases, to study its performance. With this aim,

sensitivity (SENS) and specificity (SPEC) calculations were performed. SENS refers to the percentage of cases belonging to a determinate class correctly classified and SPEC refers to the percentage of cases not belonging to a class correctly not classified in this class (Forina et al., 1991). To avoid conclusions by chance, stratified delete-a-group jackknife (DAGJK) cross-validation procedure (Kott, 2001) was followed. Stratified DAGJK randomly discards a group of samples from each class before computing the model and uses them as test samples to obtain SENS and SPEC. In this case, nine replicates were obtained by DAGJK, maintaining the ratio between training and test sets for each considered group, and mean SENS and SPEC were computed for each class (Table 5). SENS of Colima was 96%, whilst Oaxaca, Chiapas and Veracruz presented values of 83%, 67% and 63%, respectively. SPEC ranged from 91% to 96%. The overall sensitivity and specificity obtained by applying this model were 81% and 94%, respectively. In order to improve these results, a non linear pattern recognition technique, such as MLP-ANN, was applied. ANN mimic biological nervous system and can be used where other modeling techniques cannot predict complex phenomena (Zupan and Gasteiger, 1993). MLP-ANN are feed forwarded networks consisting of neurons arranged in layers (an input layer, various hidden layers and an output layer), with unidirectional connections from input to output. As LDA, ANNs learn from the data and a training set is necessary to find out the relationships between inputs and outputs, whilst a test set is used to show the prediction ability. This kind of ANN is usually trained by back propagation, minimizing the error made by the network on the training patterns. When the network starts learning the idiosyncrasies of the training cases, the model overfits the data. In order to avoid overtraining, a third set of samples (validation set) is required to estimate the generalization ability. The weights of the hidden layers are modified in order to minimize the training error and this adjustment is stopped when Table 5 SENS and SPEC (%) for LDA and MLP-ANN models. State

Chiapas Colima Oaxaca Veracruz Overall

LDA

MLP-ANN

SENS

SPEC

SENS

SPEC

67  29 96  11 83  25 63  31 81  10

94  8 93  11 96  6 91  8 94  3

100 94  12 94  17 81  24 93  10

95  8 99  4 100 97  9 98  3

R. Mun˜iz-Valencia et al. / Journal of Food Composition and Analysis 34 (2014) 7–11

the validation error begins to rise (Reed, 1993). In this case, samples were divided into training (2/4), validation (1/4) and test (1/4) sets, maintaining this proportion in each class. A network with 6 inputs (one for each raw variable selected by LDA), 5 hidden neurons and 4 outputs (one for each class) was obtained. The network was trained using a maximum of 100 epochs with learning rate of 0.1. The model was cross-validate using the same DAGJK procedure explained for LDA calculations, computing SENS and SPEC for each considered class. The samples used as training, verification and test cases were selected randomly for each replicate. As can be seen in Table 5, SENS of Chiapas, Oaxaca and Veracruz increase to 100%, 94% and 81%, respectively. In the case of Colima this parameter slightly decreases to 94%. An increase is also observed for SPEC values. Overall SENS and SPEC rise to 93% and 98%, respectively. In light of these results it can be concluded that non linear models perform better than linear ones for the discrimination of coffees from different Mexican states. 5. Conclusion Samples of Mexican coffees have been analyzed to determine the content of Ca, Cu, Fe, K, Mg, Mn, Na and Zn. The differences in the mineral composition were highlighted by means of Kruskal– Wallis multiple comparison test. Significant differences were found in the contents of Ca, Mn, Na and Zn. PCA was used to visualize the natural trends of data groups to appear separately. Some tendencies were observed in the space formed by the first three PCs. Forward stepwise LDA was used to corroborate those trends and to find out the best chemical descriptors: Ca, K, Mn, Mg, Na and Zn. The overall sensitivity and specificity of LDA were 81% and 94%, respectively. These results improved when non-linear pattern recognition techniques, such as ANN, were applied. A MLPANN model was built by using the chemical descriptors selected by LDA, allowing the differentiation of Mexican roasted coffees with 93% of prediction ability and specificity of 98%. In view of these results, Mexican coffees can be geographically differentiated according to their major elemental composition. This fact can be used in future in order to establish the differentiating characteristics of Mexican coffees growing in specific areas, as this is of great importance in the establishment of protected designations of origins. The use of major elements has the advantage of presenting very few numbers of missing data, facilitating the computation of adequate models. Acknowledgments This work has been partially supported by Project 173591 from the Government of Mexico (CONACyT) and also from University of Colima (FRABA project).

11

References Akamine, T., Otaka, A., Hokura, A., Ito, Y., Nakai, I., 2010. Determination of trace elements in coffee beans by XRF spectrometer equipped with polarization optics and its application to identification of their production area. Bunseki Kagaku 59, 863–872 (in Japanese, with English abstract). Alonso-Salces, R.M., Serra, F., Reniero, F., He´berger, K., 2009. Botanical and geographical characterization of green coffee (Coffea arabica and Coffea canephora): chemometric evaluation of phenolic and methylxanthine contents. Journal of Agricultural and Food Chemistry 57, 4224–4235. Anderson, K.A., Smith, B.W., 2002. Chemical profiling to differentiate geographic growing origins of coffee. Journal of Agricultural and Food Chemistry 50, 2068– 2075. Barham, B.L., Weber, J.G., 2012. The economic sustainability of certified coffee: recent evidence from Mexico and Peru. World Development 40, 1269–1279. Bertrand, B., Villareal, D., Laffargue, A., Posada, H., Lashermes, P., Dussert, S., 2008. Comparison of the effectiveness of fatty acids, chlorogenics acids and elements for the chemometric. Journal of Agricultural and Food Chemistry 56, 2273–2280. Carrera, F., Leon-Camacho, M., Pablos, F., Gonzalez, A.G., 1998. Authentication of green coffee varieties according to their sterolic profile. Analytica Chimica Acta 370, 131–139. Choi, M.Y., Choi, W.S., Park, J.H., Lim, J., Kwon, S.W., 2010. Determination of coffee origins by integrated metabolomic approach of combining multiple analytical data. Food Chemistry 121, 1260–1268. Cuadros, L., Garcı´a, A.M., Bosque, J.M., 1996. Statistical estimation of bS linear calibration range. Analytical Letters 29, 1231–1239. Fernandes, A.P., Santos, M.C., Lemos, S.G., Ferreira, M.M.C., Nogueira, A.R.A., Nobrega, J.A., 2005. Pattern recognition applied to mineral characterization of Brazilian coffees and sugar-cane spirits. Spectrochimica Acta Part B 60, 717– 724. Forina, M., Armanino, C., Leardi, R., Drava, G., 1991. A class modelling technique based on potential functions. Journal of Chemometrics 5, 435–453. Gonza´lez, A.G., Herrador, M.A., Asuero, A.G., 2005. Practical digest for evaluating the uncertainty of analytical assays from validation data according to the Lgc/Vam protocol. Talanta 65, 1022–1030. Historia del cafe (History of Coffee) (2012). Retrieved 02.12.12 from: http:// www.cafesdemexico.com/historiadelcafe.htm. Horwitz, W., 1982. Evaluation of analytical methods used for regulation of foods and drugs. Analytical Chemistry 54, 67A–76A. ICO-a (International Coffee Organization), 2012. Monthly Coffee Market Report. (Retrieved 02.12.12 from: http://dev.ico.org/documents/cmr-0812-e.pdf). ICO-b (International Coffee Organization), 2012. ICO Daily Indicator Prices. (Retrieved 02.12.12 from: http://www.ico.org/prices/p1.htm). International Organization of Standardization, 1994. ISO 11294. Jolliffe, I.T., 2002. Principal Components Analysis, 2nd ed. Springer, New York, USA. ˜ eiro, A., Bermejo-Barrera, P., 2007. Jurado, J.M., Martı´n, M.J., Pablos, F., Moreda-Pin Direct determination of cooper, lead and cadmium in aniseed spirits by electrothermal atomic absorption spectrometry. Food Chemistry 101, 1296–1304. Kott, P.S., 2001. The delete-a-group jackknife. Journal of Official Statistics 17, 521– 526. Massart, D.L., 1998. Handbook of Chemometrics and Qualimetrics. Part B Elsevier, Amsterdam. Reed, R., 1993. Pruning algorithms—a survey. IEEE Transactions on Neural Networks 4, 740–747. Risticevic, S., Carasek, E., Pawliszyn, J., 2008. Headspace solid-phase microextraction-gas chromatographic-time-of-flight mass spectrometric methodology for geographical origin verification of coffee. Analytical Chimica Acta 617, 72–84. Santato, A., Bertoldi, D., Perini, M., Camin, F., Larcher, R., 2012. Using elemental profiles and stable isotopes to trace the origin of green coffee beans on the global market. Journal of Mass Spectrometry 47, 1132–1140. Zupan, J., Gasteiger, J., 1993. Neural Networks for Chemists: An introduction. Weinheim, VCH.