Computers and Electronics in Agriculture 121 (2016) 313319

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture

journal homepage: www.elsevier.com/locate/compag

Original papers

Partial least square with discriminant analysis and near infrared

spectroscopy for evaluation of geographic and genotypic origin of
arabica coffee
Izabele Marquetti a, Jade Varaschim Link a,b, Andr Luis Guimares Lemes a,
Maria Brgida dos Santos Scholz c, Patrcia Valderrama a, Evandro Bona a,
a
Federal University of Technology Paran (UTFPR), P.O. Box 271, Via Rosalina Maria dos Santos 1233, CEP 87301-899 Campo Mouro, PR, Brazil
b
Federal University of Santa Catarina (UFSC), Campus Universitrio Trindade, CEP 88040-900 Florianpolis, SC, Brazil
c
Agronomic Institute of Paran (IAPAR), Rodovia Celso Garcia Cid, km 375, CEP 86047-902 Londrina, PR, Brazil

a r t i c l e i n f o a b s t r a c t

Article history: The agronomic practices and environmental conditions for coffee cultivation, such as climate, soil type
Received 2 October 2015 and altitude, promote influence in the final chemical composition of the grain. Furthermore, the genotype
Received in revised form 10 December 2015 directly influences the essential features of the beverage, increasing its aggregate price. Proof of geo-
Accepted 28 December 2015
graphic and genotypic origin of the coffee genotype must be done using reliable methods. Thus, near
Available online 14 January 2016
infrared spectroscopy (NIR) was used to analyze different coffee genotypes that were cultivated in
Brazil. Due to complexity and quantity of information within the spectra, partial least square discrimi-
Keywords:
nant analysis (PLS-DA) were applied to analyze the NIR data. The multiplicative scatter correction
Green coffee
Arabica
(MSC) and the SavitzkyGolay second-derivative were tested as preprocessing techniques to find which
Growing region one provides an appropriate identification model. The best model achieved correctly identified 94.4% of
Genotype validation samples for both geographic and genotypic origin. The results demonstrate that NIR spec-
Near infrared spectroscopy troscopy provides significant analytical data to be used in tandem with PLS-DA to distinguish green coffee
PLS-DA samples geographically and genotypically.
2016 Elsevier B.V. All rights reserved.

1. Introduction conditions affects both the chemical composition as the physico-

Coffee is one of the most consumed beverages in the world, and
its consumption has increased due to factors such as improve-
ments in the quality of the beverage, better agronomic practices,
its association with health benefits, and availability of new prod-
ucts. The quality of the beverage is associated with adequate coffee
genotype selection, which improves its flavor (Farah, 2009). Ara-
bica coffee is known for its high quality, with an intense aroma,
lower caffeine content, and a less bitter taste, and so it has higher
aggregate price (Lashermes and Anthony, 2007).
There are many arabica coffee genotypes available, most of
which were obtained by breeding that modified certain features
of the coffee. The objective of developing modern genotypes is to
obtain grains more adapted to different climatic conditions and soil
types, and also more resistant to diseases and pests, increasing pro-
ductivity and improving quality (Sera, 2001). The interaction
between the genetic variability of arabica and the cultivation
Corresponding author. Tel./fax: +55 44 3518 1477.
E-mail address: ebona@utfpr.edu.br (E. Bona).
chemical characteristics of the coffee beans (Scholz et al., 2011).
The coffee quality as a beverage depends on the chemical compo-
sition of green coffee (Ribeiro et al., 2011). Highest quality coffees
are related to the increase of sucrose, lipids, amino acids, and trigo-
nelline contents, and reduction of chlorogenic acids and caffeine
contents, which are responsible for contributing to the bitterness
of the coffee (Stalmach et al., 2006). Beans from regions and vari-
eties that are known to produce high quality beverages have a
great commercial value (Teuber, 2010). To guarantee to consumers
the geographic and genotypic origin of coffee, fast and efficient
analytic methods are required.
Different analytical techniques are often employed for coffee
analysis, including chromatographic analysis (Novaes et al.,
2015), UVVis spectroscopy (Souto et al., 2015), nuclear magnetic
resonance (Arana et al., 2015). These are slow techniques because
they require more time to prepare samples, have high costs, and
generate too much residues. To overcome these disadvantages an
alternative is employ near infrared spectroscopy (NIR) which is a
fast technique that requires minimum sample preparation, do
not destroy samples, and allows simultaneous analyses. Because

http://dx.doi.org/10.1016/j.compag.2015.12.018
0168-1699/ 2016 Elsevier B.V. All rights reserved.
314 I. Marquetti et al. / Computers and Electronics in Agriculture 121 (2016) 313319
it is a technique with a high complexity data and a large amount of
information due to overtones and combination bands (mainly
stretching and bending vibrations, from C@O, CAH, CAN, CAO,
NAH, NO2, and OAH bonds) chemometric methods are required
for spectra interpretation (Burns and Ciurczak, 2008). This tech-
nique is often used to discriminate blends of arabica and robusta
grains, obtaining satisfactory results (Esteban-Dez et al., 2007;
Santos et al., 2012; Bertone et al., 2016) or to detect defects or
adulteration in coffees (Craig et al., 2015; Winkler-Moser et al.,
2016). However, few studies used NIR spectroscopy for discrimina-
tion of arabica coffee by geographical and genotype origins. Then,
this study aimed to perform a genotypic and geographical segmen-
tation of coffees grown in Brazil by using NIR spectroscopy coupled
with partial least squares with discriminant analysis (PLS-DA).
2. Materials and methods
2.1. Coffee samples and near infrared spectra
Four Coffea arabica genotypes developed by the Agronomic
Institute of Paran (IAPAR) were evaluated in this research: IPR
99, IPR 105, IPR 106, and IA 59. The IA 59 genotype was released
in 1994, originated from the crossing between Coffea arabica vari-
eties Villa Sarchi 971/10 and Hbrido de Timor 832/2. As well
as IPR 99, is resistant to all kinds of known rust (Sera et al.,
2011). The IPR 105 genotype is derived from the Catua genotype,
while the IPR 106 genotype originated from the Icatu genotype,
and both are equally resistant to all rusts at different levels. Among
these genotypes, only IA 59 and IPR 99 genotypes are available to
farmers (Sera et al., 2010). To correlate the features of IPR 105 and
IPR 106 genotypes with IPR 99 and IA 59 (varieties already culti-
vated) could contribute for the release of new genotypes in the
market. Thus, efficient tools to help find these similarities, espe-
cially relating to chemical composition, is extremely important.
Ninety samples of coffee, carefully selected, cultivated in four
different cities were evaluated: Cornlio Procpio (CP), Paranava
(PV), Mandaguari (MD), and Londrina (LD), all in Paran State in
south of Brazil. Five sample of each genotype was used per city,
except for IA 59 samples cultivated in PV and CP, where ten sam-
ples were used for each city. The samples were harvested between
2008 and 2010. After the harvest, the beans were sent to the IAPAR
laboratories in Londrina (ParanBrazil), where they were put in
wooden boxes with a mesh bottom and moved eight times per
day until grain moisture of 1112% was achieved. Then, the sam-
ples were processed by removing the husk and the parchment.
The green beans were ground (0.5 mm) and stored in a freezer at

18 C for later analysis.
Green coffee spectra were recorded using a near infrared spec-
troscopy NIRSystem 5000-M (Foss Tecator AB, Hgans, Sweden).
Measurements were made at room temperature (23 C) in the
wavelength range 11002498 nm at 2 nm intervals. The software
WinISI III version 1.50e (Foss NIRSystems/Tecator Infrasoft Interna-
tional, LLC, Silver Spring, MD, USA) was used to acquire the spectra.
2.2. Software, preprocess and PLS-DA
All chemometric analyses for the NIR spectra were performed in
MATLAB R2007b (The MathWorks Inc., Natick, USA).
To reduce variation sources that carry no relevant information
during the multivariate calibration model, and considering scatter
effects and slope changes between samples, two spectra pre-
treatments were applied: multiplicative scatter correction (MSC)
and the SavitzkyGolay second-derivative (second order polyno-
mial with a seven points window).
MSC corrects multiplicative and additive scatter effects, which
are the result of differences in granules size, morphology, and par-
ticle orientation. It uses a linear regression of spectral variables
versus the average spectrum (Isaksson and Ns, 1988). The second
derivative using SavitzkyGolay transformation removes problems
due to slope changes between samples (Savitzky and Golay, 1964).
The PLS-DA is a supervised chemometric method applied for
classification problems developed from the two-block Partial Least
Squares (PLS) algorithms used in multivariate calibration to model
the relationship between two matrices (Barker and Rayens, 2003).
In PLS-DA, as in PLS, a linear relationship between the dependent
variable (Y) and the independent variable (X) is established. This
occur based on principal component analysis (PCA) were the matrix
X and Y are decomposed into the product of two matrices, scores
and loadings (Wold et al., 1987). In the PLS and PLS-DA methods
there are a slight rotation on the principal component axis seeking
the maximum covariance of X with Y, and now the principal compo-
nents are called latent variables (LV). The main difference between
PLS and PLS-DA is the Y matrix. The Y matrix in PLS contains the val-
ues of the property of interest, while in PLS-DA the matrix Y con-
tains information about the samples classes. For each class, the
value is assumed to be 0 or 1, depending on whether or not it
belongs to the class represented for that column (Barker and
Rayens, 2003). The modeling consists of two steps: (1) calibration,
where data characteristics are investigated to find a model for their
behavior; and (2) validation, where data that did not participate in
the calibration step are used to evaluate the model adequacy.
The threshold value for the class separation is based on Bayes
Theorem. The Bayesian threshold assumes that the predicted y val-
ues follow a distribution similar to what will be observed for future
samples. Using these estimated distributions, a threshold is selected
at the point where the two estimated distributions cross; this is the
y value at which the number of false positives and false negatives
should be minimized for future predictions. Further the model per-
formance can be evaluated using some parameters such as sensitiv-
ity and specificity. The sensitivity is the model ability to correctly
classify the samples, relating the predicted samples to being in a
class with the samples that actually are in this class. The specificity
relates the predicted samples to not being in a class with the sam-
ples that actually are not in this class (Almeida et al., 2013).
The model accuracy is evaluated using the root mean square
error of calibration (RMSEC) and prediction (RMSEP) (Valderrama
et al., 2007, 2009):
s
P 
^ p
yr 2
y
RMSEC 1
nc
m
s
P 
^ p
yr 2
y
RMSEP 2
nv
where nc is the number of samples in calibration set and nv is the
number of samples in validation set, m is the number of latent vari-
ables plus one when mean centered data is used, y^p is the predicted
value for each sample and yr is the reference value for the corre-
spondent sample.
In this work PLS-DA models were built to differentiate arabica
coffee samples by genotype and growing region. We used 72 cali-
bration samples and 18 prediction samples, all spectra data were
mean centered.
3. Results and discussion
Fig. 1 shows the NIR spectra from coffee samples, and the spec-
tra after employing the preprocess. The different harvest location is
a complicating factor during the distinction between samples due
 I. Marquetti et al. / Computers and Electronics in Agriculture 121 (2016) 313319 315

Fig. 1. Spectra of coffee samples (a) after MSC preprocess (b), after second-derivative preprocess (c) and after MSC plus second-derivative preprocess (d).

Table 1
PLS-DA results for geographic and genotypic origin. The results of the best models are marked in bold.

Geographic origin
Pre-treatment LV Class RMSEC RMSEP Calibration Prediction
Sensitivity Specificity Sensitivity Specificity
MSC + 2nd derivative 6 CP 0.1598 0.1853 1.000 1.000 1.000 1.000
PV 0.1184 0.1734 1.000 1.000 1.000 1.000
MD 0.2059 0.2186 1.000 1.000 0.750 1.000
LD 0.0919 0.1024 1.000 1.000 1.000 1.000
2nd derivative 5 CP 0.2043 0.2860 1.000 1.000 1.000 0.923
PV 0.1842 0.2459 1.000 1.000 0.800 1.000
MD 0.2109 0.1894 0.938 0.982 1.000 1.000
LD 0.1172 0.1192 1.000 1.000 1.000 1.000
MSC 4 CP 0.3360 0.3544 0.800 0.846 1.000 0.692
PV 0.3290 0.3664 0.950 0.827 0.800 0.615
MD 0.2371 0.2400 0.938 0.946 0.750 1.000
LD 0.1645 0.1734 1.000 1.000 1.000 1.000

Genotypic origin
MSC + 2nd derivative 6 IPR 105 0.2315 0.2858 1.000 0.964 1.000 0.929
IPR 106 0.2539 0.2607 0.750 0.946 0.750 1.000
IPR 99 0.2536 0.2525 1.000 0.964 1.000 1.000
IA 59 0.1439 0.1585 1.000 1.000 1.000 1.000
2nd derivative 6 IPR 105 0.2923 0.2907 0.938 0.875 0.750 0.786
IPR 106 0.2872 0.2969 0.750 0.946 0.750 1.000
IPR 99 0.2407 0.2523 1.000 0.946 1.000 1.000
IA 59 0.1727 0.2079 1.000 1.000 1.000 1.000
MSC 4 IPR 105 0.4136 0.4171 0.750 0.357 0.500 0.429
IPR 106 0.4057 0.4024 0.625 0.536 0.750 0.357
IPR 99 0.3144 0.3316 0.750 0.982 0.750 0.929
IA 59 0.3576 0.3443 0.708 0.833 0.667 1.000

to the fact that NIR spectral signature is strongly affected by the
environmental conditions (Posada et al., 2009). Nevertheless, visu-
ally it is not possible to observe significant differences between the
spectra.
3.1. Geographic origin
A PLS-DA model was developed for each spectra
preprocess Fig. 1(b)(d) in order to discriminate the coffee
samples by cultivation region. Table 1 summarize the results
obtained.
By analyzing the results from Table 1, it is possible observe that
preprocess MSC plus second-derivative promoted the best PLS-DA
model, with six latent variables that explained 95.53% of explained
variance in X and 87.80% in Y. This model was considered as the
best due the lowest values of RMSEC and RMSEP for each class
and the highest values of sensitivity and specificity. For this model,
any sample was considered outlier, since any sample presented
316 I. Marquetti et al. / Computers and Electronics in Agriculture 121 (2016) 313319

Fig. 2. Scores plot of the PLS-DA model built from spectra with MSC plus second-derivative preprocess for geographical origin. (a) LV1 against LV2 and (b) LV1 against LV3. (o)
CP, (d) PV, (+) MD, (h) LD.

Fig. 3. Loadings plot of LV 1(), 2(- - -), and 3(. . .) against wavelength (variables) for PLS-DA model built from spectra with MSC plus second-derivative preprocess for
geographical origin.

high leverage and high residuals values simultaneously (see
Electronic Supplementary Material, Fig. 1S-A).
The scores plot for the best PLS-DA model are presented in
Fig. 2, where a separation between the samples by city can be
observed, indicating that the MD samples were discriminated by
the positive part of LV1 and the negative part of LV2, while the
LD samples were discriminated by the positive part of LV1 and
LV2. The PV samples were discriminated by the negative part of
LV1 and the positive part of LV3, and the CP samples by the
negative part of LV1 and LV3.
The loadings plot in Fig. 3 shows the most intense NIR peaks,
with absolute values bigger than 0.1, which contribute to the dif-
ferentiation between classes. By analyzing this figure, it is noted
that the peaks in 1410, 1742, 1904, and 2318 nm contributed to
the MD samples classification, because they have positive loadings
for LV1 and negative for LV2. For LD samples classification, the
peaks in 1410, 1728, 1904, 2306 and 2348 nm are the most impor-
tant, because they have positive loadings for LV1 and LV2. In the PV
samples, the peaks 1436, 1880 and 2324 nm were the ones that
contributed to classification, because they have negative loadings
for LV1 and positive for LV3. For CP samples, the important peaks
are 1436, 1880, 2312 and 2350 nm, because they have negative
loadings for LV1 and LV3. The mentioned bands are related to
lipids, water, caffeine, chlorogenic acids, trigonelline, proteins,
amino acids, sugars, and carbohydrates (Ribeiro et al., 2011).
The MD samples were discriminated by peaks with bands
related to lipids, which are responsible for retaining the aroma in
coffee, while the LD samples were discriminated by peaks with
bands related to caffeine. PV and CP samples were discriminated
by peaks with bands related to chlorogenic acids. The discrimina-
tion between LD and MD samples can be related to a lower altitude
and hotter weather in Londrina compared to Mandaguari, which
 I. Marquetti et al. / Computers and Electronics in Agriculture 121 (2016) 313319 317

Fig. 4. PLS-DA classification for geographical origin. (a) Class 1 CP (o), (b) class 2 PV (d), (c) class 3 MD (+), and (d) class 4 LD (h).

Fig. 5. Scores plot of the PLS-DA model built from spectra with MSC plus second-derivative preprocess for genotype. (a) LV1 against LV2 and (b) LV1 against LV3. (r) IPR 105,
() IPR 106, () IA 59, (N) IPR 99.

makes the beans develop more quickly (Vaast et al., 2006). Similar
results were obtained in another studies in the same region
(Garrett et al., 2013). The separation between the CP and PV sam-
ples (the extremes of the longitudes analyzed) was less evident,
suggesting a similar level of adaptation in these locations. Consid-
ering that, the best quality coffees are related to an increase in
sucrose, lipids, amino acids, and trigonelline contents, and a reduc-
tion in chlorogenic acids and caffeine contents (Stalmach et al.,
2006) the MD samples probably show a high quality while LD,
PV and CP are regions that contribute to the production of coffee
grains with more bitterness. More informations about climatic
conditions and cities position can be obtained in Electronic Supple-
mentary Material, Table 1S.
The prediction results for the best PLS-DA model are presented
by class in Fig. 4. The threshold values were 0.4964, 0.5320, 0.4432,
and 0.3667 for the CP, PV, MD, and LD classes, respectively, and can
be observed by the dashed line on the plot. Only one sample, from
the MD, was classified as being from LD, thus the total correct clas-
sification of the prediction samples was 94.4%. Furthermore, the
sensitivity and specificity values obtained per class for this model
(Table 1) were close or equal to 1, indicating that the PLS-DA model
obtained can be used to classify arabica coffee by geographic
origin.
3.2. Genotypic origin
To search a discrimination between sample coffees by genotype
a PLS-DA model was developed for each spectra preprocess Fig. 1
(b)(d) and the results are shown in Table 1. The preprocess MSC
plus second-derivative had the best PLS-DA model, with six latent
318 I. Marquetti et al. / Computers and Electronics in Agriculture 121 (2016) 313319

Fig. 6. Loadings plot of LV 1(), 2(- - -), and 3(. . .) against wavelength (variables) for PLS-DA model built from spectra with MSC plus second-derivative preprocess for
genotype.

Fig. 7. PLS-DA classification for genotypic origin. (a) Class 1 IPR 105 (r), (b) class 2 IPR 106 (), (c) class 3 IPR 99 (N), and (d) class 4 IA 59 ().

variables and 92.50% of explained variance in X and 72.58% of
explained variance in Y. This model presented the lowest RMSEC
and RMSEP values for each class and the highest values of sensitiv-
ity and specificity. By regarding the residuals against leverage plot
no outliers were detected (Fig. 1S-B in the Electronic Supplemen-
tary Material).
The scores plot for this PLS-DA model are presented in Fig. 5.
A discrete separation can be observed between the samples by
genotype, indicating that IPR 105 and IPR 106 genotypes were dis-
criminated by the negative part of LV3, while the IPR 99 genotype
by the negative part of LV1 and LV2, and the IA 59 genotype by the
positive part of LV3. Probably, the spectral differences caused by
environmental interference were bigger than those attributed to
genetic differences among genotypes, which reduced the sensitiv-
ity and specificity values for genotypic classification, compared to
the geographical classification. Besides, modern genotypes, like
those analyzed in this study, have a narrow genetic base, i.e., a
similar genetic composition (Sera, 2001; Steiger et al,. 2002).
The loadings plot in Fig. 6 shows the most intense NIR peaks,
with absolute values bigger than 0.1, which contribute to the dif-
ferentiation between classes. It was noted that the peaks in 1674,
2286, 2308 and 2348 nm contributed to classification of IPR 105
and IPR 106 genotypes, due the negative loadings in LV3. For the
IPR 99 the peaks 1436, 1742, 1880, 1904, and 2322 nm were
important for this genotype distinction, because they have nega-
tive loadings for LV1 or LV2. The genotype IA 59 were classified
by the peaks 1436, 2274 and 2320 nm, with positive loadings for
LV3. As in the geographic classification, the mentioned bands are
related to lipids, water, caffeine, chlorogenic acids, trigonelline,
proteins, amino acids, sugars, and carbohydrates (Ribeiro et al.,
2011).
Based on the obtained results, it was noted that the genotypes
that are already available to farmers, IPR 99 and IA 59, presented
bands related to carbohydrates in the variables responsible for its
differentiation. Moreover, the IA 59 genotype has trigonelline
bands in the variables responsible for its differentiation. Then,
these genotypes probably produce coffee grains with high quality
when compared with IPR 105 and IPR 106 genotypes.
The threshold values for prediction of PLS-DA model by class
were 0.4859, 0.3368, 0.3662, and 0.4984 to the genotypes IPR
105, IPR 106, IPR 99, and IA 59, respectively. For genotypic origin
the PLS-DA model presented some samples uncorrected classified
 I. Marquetti et al. / Computers and Electronics in Agriculture 121 (2016) 313319
(Fig. 7). Probably, the spectral differences caused by environmental
interference were bigger than those attributed to genetic differ-
ences among genotypes, which reduced the sensitivity and speci-
ficity values for genotypic classification, compared to the
geographical classification (Table 1). Besides, modern genotypes,
like those analyzed in this study, have a narrow genetic base, i.e.,
a similar genetic composition (Sera, 2001; Steiger et al., 2002).
4. Conclusions
The PLS-DA model was able to discriminate arabica coffee
samples geographically and genotypically. The sensitivity and
specificity for calibration and prediction sets were greater in
geographic discrimination than in genotypic one, thus the
PLS-DA model for region classification is more confident.
Although the genotypes evaluated in this study having a similar
genetic composition, the NIR spectra contain important informa-
tion for discrimination by genotype. It was noted that there is a dif-
ference between coffee genotypes already available to farmers by
IAPAR (IPR 99 and IA 59) when compared to the others (IPR 105
and IPR 106). The genotype IA 59 is discriminated by the presence
of bands with trigonelline in the variables responsible for their dif-
ferentiation, suggesting a better quality coffee.
Acknowledgements
The authors thank CAPES, CNPq, Embrapa Caf and Fundao
Araucria for their financial support and scholarships.
Appendix A. Supplementary material
Supplementary data associated with this article can be found, in
the online version, at http://dx.doi.org/10.1016/j.compag.2015.12.
018.
References
Almeida, M.R., Fidelis, C.H.V., Barata, L.E.S., Poppi, R.J., 2013. Classification of
Amazonian rose wood essential oil by Raman spectroscopy and PLS-DA with
reliability estimation. Talanta 117, 305311.
Arana, V.A., Medina, J., Alarcon, R., Moreno, E., Heintz, L., Schafer, H., Wist, J., 2015.
Coffees country of origin determined by NMR: the Colombian case. Food Chem.
175, 500506.
Barker, M., Rayens, W., 2003. Partial least squares for discrimination. J. Chemom. 17,
166173.
Bertone, E., Venturello, A., Giraudo, A., Pellegrino, G., Geobaldo, F., 2016.
Simultaneous determination by NIR spectroscopy of the roasting degree and
Arabica/Robusta ratio in roasted and ground coffee. Food Control 56, 683689.
Burns, D.A., Ciurczak, E.W., 2008. Handbook of Near-Infrared Analysis, third ed.
Taylor & Francis Group, Boca Raton.
Craig, A.P., Franca, A.S., Oliveira, L.S., Irudayaraj, J., Ileleji, K., 2015. Fourier transform
infrared spectroscopy and near infrared spectroscopy for the quantification of
defects in roasted coffees. Talanta 134, 379386.
319
Esteban-Dez, I., Gonzliz-Siz, J.M., Senz-Gonzlez, C., Pizarro, C., 2007. Coffee
varietal differentiation based on near infrared spectroscopy. Talanta 71, 221
229.
Farah, A., 2009. Coffee as speciality and functional beverage. In: Paquin, P. (Ed.),
Functional and Speciality Beverage Technology. Elsevier, pp. 370395.
Garrett, R., Schmidt, E.M., Pereira, L.F.P., Kitzberger, C.S.G., Scholz, M.B.S., Eberlin, M.
N., Rezende, C.M., 2013. Discrimination of arabica coffee cultivars by
electrospray ionization Fourier transform ion cyclotron resonance mass
spectrometry and chemometrics. LWT Food Sci. Technol. 50, 496502.
Isaksson, T., Ns, T., 1988. The effect of multiplicative scatter correction (MSC) and
linearity improvement in NIR spectroscopy. Appl. Spectrosc. 42, 12731284.
Lashermes, P., Anthony, F., 2007. Coffee. In: Kole, C. (Ed.), Technical Crops: Genome
Mapping and Molecular Breeding in Plants. Springer Berlin Heidelberg, Berlin,
pp. 109118.
Novaes, F.J.M., Oigman, S.S., Souza, R.O.M.A., Rezende, C.M., Neto, F.R.A., 2015. New
approaches on the analyses of thermolabile coffee diterpenes by gas
chromatography and its relationship with cup quality. Talanta 139, 159166.
Posada, H., Ferrand, M., Davrieux, F., Lashermes, P., Bertrand, B., 2009. Stability
across environments of the coffee variety near infrared spectral signature.
Heredity 102, 113119.
Ribeiro, J.S., Ferreira, M.M.C., Salva, T.J.G., 2011. Chemometric models for the
quantitative descriptive sensory analysis of arabica coffee beverages using near
infrared spectroscopy. Talanta 83, 13521358.
Santos, J.R., Sarragua, M.C., Rangel, A.O.S.S., Lopes, J.A., 2012. Evaluation of green
coffee beans quality using near infrared spectroscopy: a quantitative approach.
Food Chem. 135, 18281835.
Savitzky, A., Golay, M.J.E., 1964. Smoothing and differentiation of data by simplified
least squares procedures. Anal. Chem. 36, 16271639.
Scholz, M.B.S., Figueiredo, V.R.G., Silva, J.V.N., Kitzberger, C.S.G., 2011.
Caractersticas fsico-qumicas de gros verdes e torrados de cultivares de
caf (Coffea arbica L.) do IAPAR. Coffee Sci. 6, 245255.
Sera, T., 2001. Coffee genetic breeding at IAPAR. Crop Breeding Appl. Biotechnol. 1,
179199.
Sera, G.H., Sera, T., Fonseca, I.C.B., Ito, D.S., 2010. Resistance to leaf rust in coffee
cultivars. Coffee Sci. 5, 5966.
Sera, T., Shigueoka, L.H., Sera, G.H., Azevedo, J.A., Carbalho, F.G., Andreazi, E., 2011.
Nova seleo da cultivar de caf IAPAR 59 com gros mais grados. In: Simpsio
de Pesquisas dos Cafs do Brasil, Arax-MG-Brazil.
Souto, U.T.C.P., Barbosa, M.F., Dantas, H.V., Pontes, A.S., Lyra, W.S., Diniz, P.H.G.D.,
Araujo, M.C.U., Silva, E.C., 2015. Identification of adulteration in ground roasted
coffees using UVeVis spectroscopy and SPA-LDA. LWT Food Sci. Technol. 63,
10371041.
Stalmach, A., Mullen, W., Nagai, C., Crozier, A., 2006. On-line HPLC analysis of the
antioxidant activity of phenolic compounds in brewed, paper-filtered coffee.
Braz. J. Plant Physiol. 18, 253262.
Steiger, D., Nagai, C., Moore, P., Morden, C., Osgood, R., Ming, R., 2002. AFLP analysis
of genetic diversity within and among Coffea arabica cultivars. Theor. Appl.
Genet. 105, 209215.
Teuber, R., 2010. Geographical indications of origin as a tool of product
differentiation: the case of coffee. J. Int. Food Agribusiness Market. 22, 277298.
Vaast, P., Bertrand, B., Perriot, J.J., Guyot, B., Gnard, M., 2006. Fruit thinning and
shade improve bean characteristics and beverage quality of coffee (Coffea
arabica L.) under optimal conditions. J. Sci. Food Agric. 86, 197204.
Valderrama, P., Braga, J.W.B., Poppi, R.J., 2007. Variable selection, outlier detection,
and figures of merit estimation in a partial least-squares regression multivariate
calibration model. A case study for the determination of quality parameters in
the alcohol industry by near-infrared spectroscopy. J. Agric. Food Chem. 55,
83318338.
Valderrama, P., Braga, J.W.B., Poppi, R.J., 2009. State of the art of figures of merit in
multivariate calibration. Quim. Nova 32, 12781287.
Winkler-Moser, J.K., Singh, M., Rennick, K.A., Bakota, E.L., Jham, G., Liu, S.X., Vaughn,
S.F., 2016. Detection of corn adulteration in Brazilian coffee (Coffea arabica) by
tocopherol profiling and near-infrared (NIR) spectroscopy. J. Agric. Food Chem.
(as soon as publishable, http://dx.doi.org/10.1021/acs.jafc.5b04777).
Wold, S., Esbensen, K., Geladi, P., 1987. Principal component analysis. Chemom.
Intell. Lab. Syst. 2, 3752.