Food Control 107 (2020) 106802

Contents lists available at ScienceDirect

Food Control
journal homepage: www.elsevier.com/locate/foodcont

Fast quantitative detection of black pepper and cumin adulterations by near- T

infrared spectroscopy and multivariate modeling
Amanda Beatriz Sales de Lima, Acsa Santos Batista, Josane Cardim de Jesus,
Jaqueline de Jesus Silva, Antônia Cardoso Mendes de Araújo, Leandro Soares Santos*
Departamento de Tecnologia Rural e Animal, Universidade Estadual do Sudoeste da Bahia, Praça Primavera - 40, Bairro Primavera, 45700-000, Itapetinga, BA, Brazil

ABSTRACT

Spices are high value-added ingredients that are part of a complex supply chain susceptible to a variety of fraudulent actions which requires fast and efficient quality
control. The study developed methods for detection of non-targeted and targeted adulteration in black pepper and cumin powder by addition of starch cassava and
corn flour (5%–50%) using NIRs associated with chemometric models. The developed non-targeted models indicated greater ease of identification of chewing in
cumin samples when compared to black pepper samples. In terms of the discrimination of the type of adulterant used the models were efficient for the two spices.
Targeted models were developed using the MLR and PLS techniques where the results presented the high predictive capacity for the different types of adulterants,
with a correlation coefficient above 0.90 and RMSE ranging from 2.2 to 7. The practical application of the models was tested in samples of powdered spices sold in
supermarkets and street markets. A high percentage of adulterated samples were observed 62% and 79% for black pepper and cumin, respectively. The results
demonstrate that adulteration in black pepper and cumin samples by addition of cassava starch and corn flour (or similar) are common, indicating the importance of
this type of investigation.

1. Introduction contamination was not accidental but rather the result of fraudulent
activities through the addition of peanut shells (FDA, 2015; Moyer,
Spices are parts of plants that, since ancient times, play a significant DeVries, & Spink, 2017; Silvis, van Ruth, van der Fels-Klerx, & Luning,
part in the diets of many as they are important ingredients in a multi- 2017). This was the most emblematic case involving cumin fraud, for
tude of foods, beverages, medicines, cosmetics, and other compositions. making use of an allergenic adulterating substance, posing a significant
Their flavor is usually stimulated by a great number of volatile organic threat to the health of entire population segments. According to Moyer
compounds, while the biological activity is related to secondary meta- et al. (2017), this example demonstrates that fraudsters may be clever
bolites like phenols and flavonoids, terpenes, oxygenated derivatized but not particularly attentive. Other examples of cumin adulteration
hydrocarbon etc. (Black, Haughey, Chevallier, Galvin-King, & Elliott, reported in the literature were the addition of almond, peanut, tree
2016; Gondaliya, Khatrani, Soni, & Baravalia, 2018; Hübert, Tiebe, & nuts, peach and cherry (Garber et al., 2016; Sicherer & Sampson, 2010)
Banach, 2016; Schweiggert, Carle, & Schieber, 2007). and poor grade fennel seeds were coated with waste marble dust and
Because of their economic importance spices have become a con- dye, and mixed in with the cumin product (John, 2012). The adulterant
stant target for adulteration. According to Black et al. (2016) and frequently used in black pepper is Papaya seeds, due to the morpho-
Galvin-King, Haughey, and Elliott (2018), this is due to the high added logical similarities, availability and low cost. This adulterant can cause
value and because these spices are part of complex and particularly liver and stomach problems, and therefore pose a health risk to the
vulnerable supply chains, generating exorbitant profits to those who unsuspecting consumer (Lakshmi et al., 2012; Vadivel et al., 2018).
practice this criminal activity. Part of the vulnerability of spices fraud is Wilde, Haughey, Galvin-King, and Elliott (2019), evaluated black
associated with the increasing consumption of these crushed and pepper with husk, spent, chili and papaya powder by means of NIR and
ground products, which makes it difficult to verify their authenticity. FTIR and observed a good classification of adulteration by PCA and
Many cases have been reported regarding fraud involving spices. In PLS-DA. However, these multivariate statistical methods are qualitative
the years 2014 and 2015 the FDA, the United States Food and Drug and do not allow the prediction of the adulterant content in new
Administration, issued 20 recall notifications of “Cumin Powder” due to samples, only the graphical classification into groups with adulteration
the presence of undeclared peanuts. The investigations showed that the bands, which limits the method.

*
Corresponding author.
E-mail addresses: [email protected] (A.B.S.d. Lima), [email protected] (A.S. Batista), [email protected] (J.C.d. Jesus),
[email protected] (J.d.J. Silva), [email protected] (A.C.M.d. Araújo), [email protected] (L.S. Santos).

https://doi.org/10.1016/j.foodcont.2019.106802
Received 15 May 2019; Received in revised form 29 July 2019; Accepted 1 August 2019
Available online 01 August 2019
0956-7135/ © 2019 Elsevier Ltd. All rights reserved.
A.B.S.d. Lima, et al. Food Control 107 (2020) 106802

Fig. 1. Workflow for the data analysis and development models.

In addition to the adulterants mentioned above, starch has been them into useable analytical parameters (McGrath et al., 2018).
identified as an adulterant of condiments such as garlic, ginger, and The objective of the present study was to develop a method for the
powdered onion, in order to obtain economic benefit by diluting a detection of adulteration in black pepper and cumin powder through
valuable ingredient with a cheaper one (Lee et al., 2014, 2015; Lohumi the addition of simple amylaceous (cassava starch) and complex sources
et al., 2014; Lohumi, Lee, & Cho, 2015). The addition of this component (corn flour) using NIRs associated with chemometric models.
in genuine (pure) condiments constitutes food fraud. The use of amy-
laceous raw materials, such as corn and cassava, as adulterants occur
due to the great availability in nature, being an option of easy access, 2. Material and methods
low cost and inert regarding color, smell and flavor. In addition, in
some spices, starch is the main component, as is the case of black 2.1. Sample collection and preparation
pepper, where starch may represent up to 50% of dry weight (Bhat &
Tharanathan, 1983), which makes it difficult to adopt robust strategies Fig. 1 shows the experimental design, a total of 90 samples of black
to detect fraudulent adulterations. pepper (BP) and 40 of cumin (CU) were investigated in this study. The
In the detection of adulteration of herbs and spices, a number of samples were collected in February 2018 (nBP = 11 and nCU = 10)
spectroscopic methods continue to be developed. Among these, vibra- and June de 2019 (nBP = 69 and nCU = 30) in the states of Bahia
tional spectroscopies (as Near-Infrared Spectroscopy, NIRS), along with (nBP = 39 and nCU = 18), Espirito Santo (nBP = 31) and Alagoas
chemometrics, have become well known as rapid, low-cost, simulta- (nBP = 20 and nCU = 12), Brazil. As for the source of the samples,
neous analyses or characterization of multiple components, non-de- were collected directly from producers (nBP = 30), spice houses
structive, fingerprinting techniques and are valuable screening tools in (nBP = 14 and nCU = 16), street market (nBP = 30 and nCU = 20)
the detection of adulteration/authentication in the food industry (Ellis and the others in supermarkets (nBP = 14 and nCU = 6). Of the black
et al., 2012; Galvin-King et al., 2018; Magwaza, Messo, Laurie, Laing, & pepper samples used in this study, eleven were submitted to drying in
Shimelis, 2016). In spectroscopy, following the collection of scans or the ground, eleven in the conventional dryer and eight in a solar dryer.
“fingerprints”, results are evaluated using chemometric models as the For the other samples of pepper and cumin, the drying process is un-
raw spectra are too complicated to process visually. Chemometric known. Samples were purchased, both whole grains, so that it was
models extract the important information that distinguishes different possible to attest to their authenticity. From the data set concerning the
clusters and therefore ignores redundant data and simplify this process. samples of black pepper 30 were collected directly from bona fide
Chemometrics uses mathematical and statistical modeling to recognize producers with a known origin. These samples were used as a reference.
patterns and relationships within highly complex data and translate After obtaining the NIR spectra the total set of samples were submitted
to Principal Component Analysis with the objective of verifying the

2
A.B.S.d. Lima, et al. Food Control 107 (2020) 106802

similarity of the spectra of the reference samples with the others. 2.4. Exploratory analysis - principal components analysis
Samples with dispersion different from the references were discarded as
susceptible to fraud and the like were considered authentic. The NIR Principal component analysis (PCA) was used to identify the inter-
spectra of the non-spiked samples were compared to the NIR spectra relationships between the samples and possible clusters, select the
obtained for genuine black pepper samples described in the works of suitable experimental data for model construction and to identify and
Vadivel et al. (2018) and Wilde et al. (2019). The cumin samples had eliminate the outliers. PCA is a linear pattern recognition technique that
their authenticity attested by means of the visual analysis considering allows extracting the most relevant information from an X dataset by
the following characteristics: seeds are small elongated, oval and attain projecting it into latent variables (LVs), by a linear combination of the
light yellowish to grayish-brown in color. Cremocarp oblong-ellipsoid, original variables, thus reducing the dimensionality of multivariate
slightly laterally compressed, 4–6 mm in length and 1.5–2.5 mm in data to n principal components non-correlated. The objective is to
diameter. Both ends slightly acute, a small fruit stalk at the base. maximize the explained variance in the first latent variables, thus, each
Mericarp oblong-ellipsoid, the dorsal side consists of three blunt and new orthogonal space accounts for less explanatory power than the
prominent ridges, with secondary rib between every two ridges; com- previous (Genisheva et al., 2018; Sampaio et al., 2017). The number of
missural surface with a prominent longitudinal rib (Jain, Modi, & principal components was chosen by the evaluation of cumulative
Pithadiya, 2009; Ma, Mao, Zhou, Li, & Li, 2015). At the same time, variance, where the first PCs were chosen, accumulating a percentage
Principal Component Analysis was used to identify suspect samples that of variance higher than 70% and an eigenvalue greater than one. The
require an exclusion or further investigation. After collection, the variables that showed a low correlation with the principal components
samples were ground in a Ball Mill (MA 350, MARCONI) for 3 min. of higher variance and a similar correlation between the main compo-
Then, the black pepper and cumin powder samples were adulterated nents were discarded.
with cassava starch and corn flour, respectively. The addition was
carried out in order to obtain the following concentrations of 0%, 5%,
10%, 20%, 30%, 40% and 50% (g adulterant/100 g black pepper or 2.5. Targeted modeling
cumin) adulterants. Because of the spiked samples, 630 and 280 ex-
perimental units were obtained for black pepper and cumin, respec- In order to determine the level of adulterant in the black pepper and
tively. With the objective of testing the targeted and non-targeted va- cumin liners used the RML and PLS techniques.
lidation models that will be developed in this work in real samples, 19
samples of cumin and 13 black pepper powder of different brands were
2.5.1. Multiple Linear Regression
collected.
The concentrations of adulterants were correlated with the spectral
information using the major peaks by the MLR regression method. The
2.2. Near-infrared spectroscopy (NIRS) analysis was performed initially to remove samples considered as out-
liers. The data set for each adulterant percentage (14 groups for each
Spectra were collected in a NIR spectrometer (SpectraStar 2500XL, spice, being 7 for each adulterant) was then divided into two subsets
Unity Scientific, Brookfield, CT, USA) equipped with a tungsten halogen using the Kennard-Stone (KS) algorithm (Kennard & Stone, 1969): i) the
lamp as light source, an indium–gallium–arsenide (InGaAs) detector. calibration set with 70% of the samples, from which it was possible to
The signals were generated in reflectance (%R) mode and transformed establish the regression functions; and (ii) the validation set with the
into absorbance by using log 1/R. Samples were poured into the sample remaining 30% of the samples, which were used to verify the gen-
cup and scanned over the range of 1100–2500 nm at 1 nm intervals. eralization capacity of functions formed in the training.
Each point is the average score of 64 scans with 1400 points. During Three optimization methods were tested: Backward, Forward and
collection, the temperature was maintained at around 25 °C in the la- Stepwise. In the Backward technique all the predictors were included at
boratory. The Unity InfoStar V3.11.3 software was used for spectro- one time in the regression equation generated and then were with-
meter configuration, control, and data acquisition. A total of 912 NIR drawn one by one, in order of higher probability values (p > 0.05),
spectra were acquired, 910 being pure and adulterated spice spectra until the best predictors were identified, that is, those that presented
and 2 spectra of pure adulterant samples. significance (p < 0.05), in addition to low inflation of the variance. In
Forward, the equation started with only one predictor and each new
predictor that presented significance (p < 0.05) was added in-
2.3. Spectral data acquisition dividually in the equation until all the predictors added were able to
explain the phenomenon under study. The Stepwise technique is char-
Prior to the development of multivariate analyses, the spectral data acterized by including predictors in blocks until the best-generated
are pre-processed. The data were organized in a four matrix format X model fits the data (Tabachnick & Fidell, 2006).
(70 × 1400) and W (70 × 1400) where each row corresponds to cumin The MLR model was tested according to Equation (2).
and black pepper, genuine and fraudulent, sample and the columns
correspond to the absorbance (log 1/R) values. In the Principal
y= 0 + 1*x1 + …+ n*x n + (2)
Component (PCA) and Multiple Linear Regression (MLR) Analysis, the
Where: y = response or dependent variable (percentage of adultera-
variables used were the values of maximum absorbance associated with
tion); 0 = constant (intercept); x1 … xn = explanatory or independent
the peaks in the fraudulent, genuine and adulterating sample spectrum.
variables (absorbances); 1 … n = model parameters (regression
These variables were standardized in order to create a new variable
coefficients); Ɛ = model errors.
(Z) of mean equivalent to zero and standard deviation equal to one, to
Among the three techniques tested for each adulterant type in each
eliminate the differences between the measures of the variables to be
spice, the performance of the final MLR model was evaluated according
studied (Equation (1)).
to the RMSE, the determination coefficient (R), the ratio of performance
Z = (X µ )/ (1) to deviation (RPD) and range error ratio (RER). RMSE is defined as:

n
Where, Z = Standardized variable; X = Original variable; μ = Mean, RMSE = (yi yˆi )²/ n
and σ = Standard deviation. i=1 (3)
All the multivariate statistical analyses were performed in the
Statistical Analysis System (SAS) software, Student version. the coefficient of correlation is:

3
A.B.S.d. Lima, et al. Food Control 107 (2020) 106802

n n Classification Model was developed with the objective of distinguishing

R= 1 [ (yi yˆi )2 / (yi yˆ) 2] between genuine and fraudulent samples with cassava starch or corn
(4)
flour providing information on the type of adulterant used, when pos-
i=1 i=1

the ratio of performance to deviation is: sible. The samples of cumin and black pepper were divided into three
n n
groups, the first consisting of samples of genuine spices (GS Group). The
RPD = DP /(RMSEC or RMSEV ) = ( (yi y )2/(n 1) )/( (yi yˆi )2/(n 1) ) second compound samples adulterated with cassava starch (CS Group)
i=1 i=1 and finally the group constituted by samples adulterated with corn flour
(5) (Group CF). For each of these groups, the data set was divided into two
subsets, as described in item 2.5.2.
moreover, the range error ratio is:
The chemometric analysis for the development of the non-targeted
n
models was performed using the SIMCA 16 software (Umetrics,
RER = Reference amplitude/(RMSEC or RMSEV ) = (ymáx ymin )/( (yi yˆi )2/(n 1) )
i=1 Sweden). In order to obtain the validation models binary and multiclass
(6) was used a single-class classifier (OCC) technique SIMCA (soft in-
dependent modeling of class analogy) and discriminant PLS-DA (partial
where n is the number of samples in the test set, yi is the experimentally least square discriminant analysis) and OPLS-DA (partial orthogonal
measured reference result for sample i and ŷi are the estimated results of least square discriminant analysis).
the model for the corresponding test sample i. (Equation (3)). The The central feature of the SIMCA method is the application of PCA
correlation coefficient (R) between the predicted and the measured to the sample category studied, generally after-class autoscaling or
values were calculated for both the calibration and the validation test centering. SIMCA models are defined by the range of the sample scores
sets with Equation (4), where y is the mean of the reference measure- on a selected number of low-order main components (PCs) referred to
ment results for all samples in the calibration and test set. RMSEC and as the SIMCA inner space. The ideal quantity of principal components is
RMSEV are equivalent to the mean square error for calibration and determined independently for each class and the final model is obtained
validation data, respectively (Equations (5) and (6)). For Equation (6), by defining the boundary between each modeled PCA. As a con-
the amplitude equals the maximum and minimum values of reference sequence, a hyperbox with as many dimensions as the number of CPs
fraud percentages. The best combination of spectral regions and the most suitable for each class is obtained. Therefore, we have models
preprocessing techniques were selected by picking the MLR model with corresponding to rectangles (two PCs), parallelepipeds (three PCs), or
a low RMSE, a high R and a low number of variables covering enough hyper-parallelepiped (more than three PCs). At the second step, the PCA
data variance. Also, a combination of RPD and RER values that will be results are used for calculating two relevant distances for each object
discussed later. i = 1, …, I of the training set (Fidelis et al., 2017; Oliveri & Downey,
Graphs of the predicted and actual validation data of all MRL 2012; Oliveri, Casale, Casolino, Baldo, & Grifi, 2011).
models were generated to evaluate the fit of the models. Regression An important group of pattern-recognition tools is represented by
coefficients graphs in relation to wavelength were generated to eval- the discriminant-classification techniques PLS-DA and OPLS-DA. PLS-
uate which are the main predictable variables of the models. DA is based on a conventional PLS regression method, where the (I × J)
matrix X is a predictor matrix, and the (I × K) response matrix Y
2.5.2. Partial Least Squares comprises categorical (dummy) variables that describe class member-
Partial least squares (PLS) is the most popular multivariate tool to ships. K is equal to the number of classes. If only two classes are con-
develop a calibration model in order to relate the information of in- sidered, matrix Y is reduced to an (I × 1) vector y. When PLS regression
terest of a sample to its spectrum (Chen, Tan, & Lin, 2018). The ma- is developed, the response value Ypred is predicted for a new sample.
trices containing the data provided by the NIR spectra, denominated by The decision is based on the comparison of Ypred with given categorical
X, and the vector Y containing the adulterant concentrations were variables in Y. The sample is attributed to the class, which has minimal
employed to build the regression model. The PLS regression was per- distance between Y and Ypred (Rodionova, Titova, & Pomerantsev,
formed after outlier identification. 2016). The main benefit in interpretation using OPLS-DA compared to
Data were randomized and separated into training and validation PLS-DA thus lies in the ability of OPLS-DA to separate predictive from
data sets for each fraud test, as described in 2.5. The number of latent non-predictive (orthogonal) variation (Bylesjo et al., 2006).
variables (LV) in the models was selected according to the method Models performances were evaluated by the ability to detect the
proposed in the software. In fact, the T2 test was used to determine if presence and the type of adulterant in spices samples through the
two PRESS values were significantly different. If there were no sig- sensitivity (Equation (7)) and specificity (Equation (8)). The sensitivity
nificant difference, more parsimonious models were chosen and only is the percentage of the correctly detected pure black pepper and cumin
those having fewer factors than the minimum PRESS model were in relation to all pure black pepper samples in the test set. The speci-
compared. After selecting the number of latent variables, the generated ficity is the percentage of detected adulterated samples in relation to
PLS model was used to predict the concentration of adulterants in black the adulterated samples in the test set.
pepper and cumin in the validation data. The values of the coefficient of
correlation (R) and root-mean-square error (RMSE) were obtained for Sensitivity (%) = (TP /(TP + FN )) × 100 (7)
calibration and validation data.
Specificity (%) = (TN /(TN + FN )) × 100 (8)

2.6. Non-targeted modeling Where, TP = true positive (correctly identified samples); FN = false
negative (incorrectly rejected samples); TN = true negative; and
In order to fit the idea of non-targeted Analysis, cassava starch and FP = false negative.
corn flour were used as adulterants in the spices under study because of
the possibility of chemically representing several groups of foods of 2.7. Collection of commercial samples
simpler composition (rich in starch only) and another more complex
than they would have in addition to starch, simpler sugars, proteins, fat, In order to verify the practical application of the non-target and
and carotenoids. Initially, a non-targeted validation (Binary) model was target models, commercial samples were collected in street Market and
developed that indicates the authenticity of the sample through the supermarkets of powdered spices in June 2019, with thirteen samples
contrast between the genuine and adulterated samples, without dis- of black pepper and nineteen samples of cumin. These samples were
tinguishing the source of adulteration. In a second moment, a Multiclass submitted to the NIR test as described in item 2.3 and their spectra were

4
A.B.S.d. Lima, et al. Food Control 107 (2020) 106802

Fig. 2. Mean absorbance spectrum of black pepper and cumin pures and adulterates with cassava starch (A and C) and corn flour (B and D) samples, respectively.
Black line: Pure Spice; Red line: 5% adulterant; Green line: 10% adulterant; Yellow line: 20% adulterant; Blue line: 30% of adulterant; Pink line: 40% adulterant; Sky
blue line: 50% adulterant; and Gray line: Pure Adulterant.

treated as described in item 2.4. Then the obtained absorbances were relation to the cassava starch, composed mostly of carbohydrates
submitted to the non-target binary models in order to determine the (Oladunmoye, Aworh, Maziya-Dixon, Erukainure, & Elemo, 2014;
presence or not of adulteration. When it was found to be a non-genuine Qamar, Aslam, Huyop, & Javed, 2017). This behavior is less evident in
sample, the data were submitted to the multiclass models to define the the cumin samples since in the range of 1100–1400 nm it is possible to
type of adulterant used. Finally, adulterated samples had their levels visualize differences in the absorbance intensity of the samples with the
determined using a PLS or MLR regression model. increase of the corn flour concentration.
The workflow for the data analysis and model development is However, there are no specific functional groups that allow to
summarised in Fig. 1. qualitatively differentiate samples of black pepper and cumin from
adulterants used without the need for multivariate techniques. Table 1
shows the main peaks associated with pure adulterants and spices.
3. Results and discussion
The difference between the spectra is due to the absorbances asso-
ciated with the groups, which may increase or decrease according to the
3.1. NIR spectra of black pepper, cumin, and adulterants
variation of the composition and chemical complexity of the samples.
This spectral behavior can be discussed and characterized by the ab-
The analysis of the spectra obtained did not show random noise that
sorption bands of water, carbohydrates, proteins, and lipids. According
could be removed by smoothing techniques such as Savitzky-Golay
smoothing. However, we observed systematic variations with slope in
the baseline (bias) “going up a hill” as the number of waves decreases Table 1
Main peaks (nm) associated with pure adulterants and spices.
that could be corrected by means of applying the method of the second
derivative, but they have the drawback of decreasing the relation Cassava Starch Corn Flour Black Pepper Cumin
signal/noise, which occurred in a very drastic way for this study. The
1100 1100 1100 1100
spectra presented in Fig. 2 were obtained from the mean absorbance of
1203 1202 1194 1192
the NIR of each adulterant for the different concentrations and type of 1340 1340 1340 1340
adulterant used. 1465 1466 1466 1393
In Fig. 2A and C, the absorbances of the pure spices are higher than 1580 1580 1705 1464
1706 1706 1754 1725
the absorbances of the cassava starch. This makes it possible to obtain
1776 1776 1933 1759
spectra that allow the visualization of the effect of adding the adul- 1933 1932 2108 1935
terant by reducing the absorbance of the spectra when the addition of 2102 2101 2287 2104
cassava starch to the cumin and black pepper is carried out. The spectra 2284 2286 2312 2169
of spices with the addition of corn flour (2B and 2D) show peaks with 2318 2316 2348 2311
2500 2500 2500 2346
overlaps that may hinder visual differentiation. This is probably the
– – – 2391
most complex effect on the chemical composition of corn flour, that – – – 2500
features carbohydrates, proteins, and lipids in its composition in

5
A.B.S.d. Lima, et al. Food Control 107 (2020) 106802

Fig. 3. PCA score plots of black pepper and cumin,

adulterated with cassava starch (A and C) and corn
flour (B and D), respectively. Black symbol: Pure
Spice; Red symbol: 5% adulterant; Green symbol:
10% adulterant; Yellow symbol: 20% adulterant;
Blue symbol: 30% of adulterant; Pink symbol: 40%
adulterant; Sky blue symbol: 50% adulterant.

to Chakravartula, Cevoli, Balestra, Fabbria, and Rosa (2019), adulterated corn flour, the separation was similar to that of the black
(Mehrotra, 2006) and Ramos-Diaz, Rinnan, and Jouppila (2019), at pepper with this same adulterant, so only the adulterated samples with
around 1450 and 1940 nm, the resonance bands of water and starch- 30%, 40%, and 50% became more compacted.
related O–H bonds, already at about 1500–1570 nm and The lack of resolution observed in the formation of some groups
2050–2070 nm, may be related to the presence of N–H bonds. While the may be associated to the fact that PCA is an unsupervised technique
absorption peaks at 1730, 1770 and 2310 nm correspond to the lipids used to generate indexes and grouping of individuals according to their
and can be attributed to the first harmonic of the C–H stretch and the variances (Hongyu, Sandanielo, & Oliveira Junior, 2015). On the other
C–H2 group. According to Cevoli, Gianotti, Troncoso, and Fabbri hand, Multiple Linear Regression (MLR) as well as Partial Least Squares
(2015), the 2070–2100 nm region corresponds to the combined bands Regression (PLS) can use smaller sources of variation to promote group
of O–H (starch) bonds, whereas the 2280–2340 nm and 2180–2200 nm separation in addition to predicting percentages of adulteration for each
regions correspond to the C–H bonds (starch, proteins, and lipids) and fraud by means of multivariate calibration.
C=O (protein) bonds, respectively.
3.3. Targeted modeling
3.2. Exploratory analysis - principal components analysis
For the MLR, the outliers were removed and different mathematical
The absorbances associated with these wavelengths were used as methods of adjustment optimization were tested, using coefficient
variables in the identification and quantification of frauds. This was maximization and correlation (R) and root mean square error (RMSE)
possible due to the treatment of the variables with the multivariate as the criterions. The calibration equations were obtained, and the ca-
statistics, PCA and MLR. The graphics with scores of principal compo- libration and validation results of the model are expressed in Table 2.
nents for each fraud are shown in Fig. 3. It took only two principal The method that best fit the cumin samples for setting optimization
components to describe more than 90% of the total variability of the of the MLR model was Backward and is also the method that used the
data for both frauds in both spices, being effective in reducing the di- least number of variables for the calibration and prediction of fraud
mensionality of the data. content. For the black pepper samples, the Forward optimization
The formation of groups according to degrees of fraud for black method was that which produced the highest correlation coefficient and
pepper with cassava starch, as: without adulteration, with samples of the lowest RMSE among the three methods tested. It was also the
pure black pepper; low degree of adulteration with 5% and 10% of method that used the largest number of variables for calibration of the
adulterant; medium degree of adulteration with 20%; and high degree regression. This is probably due to the high content of starch present in
of adulteration, with 30%, 40% and 50% of adulterant, is observed the pure pepper composition (Bhat & Tharanathan, 1983) which may
through the PCA. However, for the black pepper adulterated with corn make it difficult to detect this type of fraud, requiring a greater number
flour, the separation was more difficult. Only two groups were formed; of variables.
one group was composed of pure and adulterated samples with 5%, Analyzing the calibration and validation of MLR, satisfactory ad-
10%, and 20%; another group was formed with samples of 40% and justments were observed with a correlation coefficient between the
50% of adulteration, while the samples with 30% of this fraud were percentage of actual and predicted adulterant higher than 0.97 and
dispersed in the first and second group. For samples of cumin adulter- 0.91 for samples of cumin and black pepper, respectively. In general,
ated with cassava starch, it was possible to verify a greater dispersion of the estimated error (RMSE) was higher for the black pepper samples
the samples with 5% of adulteration. While the pure cumin samples and than for the cumin samples. The RMSE for the validation data of the
those adulterated with 10% and 20% became more compacted and adjusted models indicates that, on average, there may be an error of up
formed patterns. The samples with 30%, 40%, and 50% showed the to 2.63% and 7.03% in the prediction of the percentage of adulterants
same pattern and therefore formed the same group. However, for the in cumin and black pepper, respectively. These values refer to a low

6
A.B.S.d. Lima, et al. Food Control 107 (2020) 106802

Table 2
MLR calibration parameters of the NIR spectra to predict the adulterated content of cumin and black pepper.
Treatment Method Number of variables Calibration Validation

R RMSE RPD RER R RMSE RPD RER

PCS Forward 14 0.9731 4.2031 4.1047 11.8959 0.9546 5.2266 3.2916 9.5393
PCF Forward 13 0.9730 4.0536 4.2691 12.3347 0.9174 7.0295 2.4551 7.0937
CCS Backward 14 0.9903 2.5360 6.6840 19.7211 0.9784 3.7021 4.5464 13.4141
CCF Backward 10 0.9922 2.1589 7.9117 22.4862 0.9901 2.6256 6.3117 17.9388

PCS: Black pepper adulterated with cassava starch; PCF: Black pepper with corn flour; CCS: Cumin adulterated with cassava starch; CCF: Cumin adulterated with corn
flour.

Fig. 4. Predicted and reference values of the adulteration percentage for validation data for black pepper and cumin MLR, adulterated with cassava starch (A and C)
and corn flour (B and D), respectively.

associated error for prediction of single or complex starch adulterants in pepper scams and highly predictable for cumin scams. According to
cumin samples. For the black pepper samples, the error associated with Williams and Sobering (1993) models with RER < 3 have little pre-
the prediction of the concentrations of these two adulterants can be dictive ability, 3 < RER < 10 have moderate practical utility, and RER
considered moderate. values > 10 indicate good practical utility.
Standard error of performance to standard deviation (RPD) is used The dispersion plots shown in Fig. 4 present the correlation between
as a measure of the ability of a NIRS model to predict a constituent, the experimental adulterant content and the adulterant content pre-
where: excellent models RPD > 2, fair models 1.4 < RPD < 2 and dicted by the model chosen for each spice. Fig. 4 shows a small and
non-reliable models RPD < 1.4 (Chang, Laird, Mausbach, & Hurburgh, uniform dispersion of data across all levels of fraud, indicating the
2001; Williams & Sobering, 1993). According to this classification, the quality of the estimated models. Higher dispersion was observed in
MRL models for both frauds and adulterants are excellent models. None adulterated pepper samples.
of the mentioned thresholds have a statistical basis in the literature Fig. 5 is the regression coefficients as a function of the wavelengths
(Ncama, Magwaza LMditshwa, & Z.Tesfay, 2019). of the spectra for the MRL models.
The range error ratio (RER) shows that the predicted models de- From Fig. 5 we can observe the regression coefficients associated
veloped in this work are of moderate practical utility for the two black with the significant vibrational modes (p < 0.05) for the MRL model.

7
A.B.S.d. Lima, et al. Food Control 107 (2020) 106802

Fig. 5. Regression coefficients as a function of the wavelength of the RLM model constructed to predict the concentration of cassava starch (A and C) and maize flour
(B and D) in black pepper and cumin, respectively.

Table 3
PLS regression calibration parameters of the NIR spectra to predict the adulterating content of cumin and black pepper.
Treatment Latents Variables Calibration Validation

R RMSE RPD RER R RMSE RPD RER

PCS 10 0.9739 3.9327 4.4051 12.6987 0.9541 5.2258 3.2921 9.5408

PCF 15 0.9834 3.1367 5.5127 15.9220 0.9581 4.9980 3.4531 9.9770
CCS 12 0.9915 2.2574 7.6802 22.0892 0.9908 2.4484 7.0130 20.2880
CCF 10 0.9907 2.0916 7.3343 21.0712 0.9925 2.2381 7.6144 22.2073

PCS: Black pepper adulterated with cassava starch; PCF: Black pepper with corn flour; CCS: Cumin adulterated with cassava starch; CCF: Cumin adulterated with corn
flour.

Based on these coefficients it is possible to indicate the most important Table 3 shows that 10 to 15 latent variables were required. The PLS
variables for the prediction of the models and also to evaluate if they model uses the information of the data matrix X and the matrix of
are really relevant for the study (Pasquini, 2018). The main peaks re- concentration Y obtaining new variables called latent variables (LV).
sponsible for the prediction of these models are similar and are asso- Only a few of these LVs explain the variation of the data. The amount of
ciated with starch (2070–2100 nm), lipid (1770 and 2310 nm) and LV is determined by the T2 test, which considers a T2 greater than 10%
protein (1500–1570 nm) binding (Chakravartula et al., 2019; Mehrotra, (α > 0.1) as a criterion for choosing the minimum number of these
2006; Ramos-Diaz et al., 2019). This points out that the models sa- variables. Therefore, the lower the number of significant LV the greater
tisfactorily associate the composition variations of the samples to gen- the ease of the model to predict the values with a low associated error
erate the models. and the simpler the model becomes.
In order to reduce the dispersion still existing between the predicted The model generated by PLS for black pepper samples required a
and experimental data, a full-spectrum method was used to calibrate a similar number of LV for a better fit when comparing the cumin samples
model using the absorbances associated to all the wave numbers mea- for the two types of fraud. The PLS model showed relatively higher R
sured in the NIR and the Regression by Partial Least Squares (PLS). This and lower RMSE overall for the samples tested when compared to the
method is not based on the hypothesis that the variables are un- MLR for the two spices and this increase is most evident in samples of
correlated and does not require the residues to follow a normal dis- black pepper. The RPD and RER values were considerably higher for the
tribution. The PLS method uses the factors obtained in order to max- PLS predicted models than the MRL prediction models. This indicates
imize the covariance between the independent variables and the that the technique was efficient to increase the robustness of the
dependent variables (Garcia & Filzmoser, 2017). The evaluation para- method for all treatments.
meters of this adjustment for the level of adulterants in black pepper The correlation between the actual and predicted adulteration
and cumin were the same as those adopted for MLR and are expressed content by the PLS model for validation samples is shown in Fig. 6. This
in Table 3. figure shows a small and uniform data dispersion for all levels of fraud.

8
A.B.S.d. Lima, et al. Food Control 107 (2020) 106802

Fig. 6. Predicted and reference values of the adulteration percentage for validation data for PLS regression of red pepper and cumin, adulterated with cassava starch
(A and C) and corn flour (B and D), respectively.

A better fit for validation of the model with black pepper samples, when Regarding the specificity of the SIMCA method, the binary and multi-
compared to MLR, is evidenced graphically by the lower dispersion of class models for the cumin presented the best values pointing to the
the data. However, for the prediction of the adulteration content in correct classification capacity of the fraudulent samples even when
pure black pepper samples for the two adulterants it was more difficult specifying the types of adulterants used.
to obtain negative values, from which probably most of the error The PLS-DA and OPLS-DA discrimination methods were notable for
(RMSE) associated with the prediction occurs. their high discrimination capacity of adulterated (specificity) samples
The regression coefficients as a function of the wavelengths of the and low ability to properly classify cumin or black pepper genuine
spectra for the PLS model are shown in Fig. 7. samples. The specificity values for these methods ranged from 90% to
For PLS models, it is generally observed that the most relevant 100%. In terms of prediction ability of adulterated samples the OPLS-
predictive variables are distributed over the entire spectrum. Where the DA method is nearly identical to PLS-DA. According to Bylesjo et al.
vibrational modes associated with the OH and CH functional groups, (2006), the main benefit of OPLS-DA modeling lies in the ease of in-
mainly related to the most important starch and lipid components. terpretation, especially in the multiclass case. This is achieved by the
The model generated by PLS was more efficient to predict the two separate modeling of predictive and class-related variation in the X-
adulterants for black pepper, considering R, RMSE, RPD, and RER. For matrix through the identification of Y-orthogonal variation. This can
the cumin samples, the two models were efficient. However, due to the help ensure both the stability of the classifier and the understanding of
greater simplicity of the MLR model that uses only the main peaks and the system studied.
the lower RMSE, this model may be more indicated. In general, the results of the application of SIMCA in binary and
multiclass models demonstrate that this method has the potential to
3.4. Non-targeted modeling reveal possible adulteration in powdered cumin samples. In terms of
black pepper adulteration, the results point to the need to use the
Table 4 presents the results obtained from the SIMCA, PLS-DA and SIMCA classification techniques and OPLS-DA or PLS-DA discrimination
OPLS-DA methods. Among these methods, SIMCA was the only one that to obtain a method that presents high values of sensitivity and speci-
presented high sensitivity indicating the ability to correctly classify ficity. Thus, the following criteria would be adopted for the binary
genuine black pepper and cumin samples. These results were better for classification model of a new sample in one of the groups: i) Identify as
the binary models with 100% and 92% correct classification rate for the genuine the samples that were thus classified initially in the SIMCA and
unadulterated samples of black pepper and cumin, respectively. then in the OPLS-DA; ii) Identify as adulterated sample those that were

9
A.B.S.d. Lima, et al. Food Control 107 (2020) 106802

Fig. 7. Regression coefficients as a function of the wavelength of the PLS model constructed to predict the concentration of cassava starch (A and C) and maize flour
(B and D) in black pepper and cumin, respectively.

Table 4 Table 5
Classification results of SIMCA, PLS-DA and OPLS-DA models used in the Test of the non-targed and targed models in commercial samples of cumin and
classification of Black pepper and Cumin according to adulterant. black pepper powder.
Sample Model Merit Figure SIMCA PLS-DA OPLS-DA Spice Class Percentage of the Level Adulteration
sample (%)
Black Pepper Genuine Binary Sensitivity 100% 10% 25%
Black Pepper Binary Specificity 76% 99% 100% Black Pepper Genuine 38% ─
Adulterated Adulterated 62% ─
Cumin Genuine Binary Sensitivity 92% 16% 32% Cassava Starch 50% 2.9 a 9.17
Cumin Adulterated Binary Specificity 96% 99% 100% Corn Flour 38% 7.8 a 53.9
Black Pepper Genuine Multiclass Sensitivity 88% 24% 36% No Class 12% ─
Cassava Starch Multiclass Specificity 83% 92% 90% Cumin Genuine 21% ─
Corn Flour Multiclass Specificity 57% 98% 99% Adulterated 79% ─
Cumin Genuine Multiclass Sensitivity 75% 33% 42% Cassava Starch 40% 4.3 a 26.4
Cassava Starch Multiclass Specificity 94% 98% 98% Corn Flour 60% 30.0 a 64.7
Corn Flour Multiclass Specificity 93% 92% 93% No Class 0% ─

thus classified initially in SIMCA and OPLS-DA; iii) Samples classified in the cumin and black pepper samples are dispersed throughout the spectrum.
SIMCA as genuine and in OPLS-DA as adulterated are inconclusive. This Particularly in the water resonance bands and the O–H bonds related to the
may be a false positive or false negative; (iii) Samples identified as starch, the presence of N–H bonds and the lipids attributed to the first
adulterated should be subjected to priority classification models in harmonic of the CH stretch and the C–H2 group.
order to verify the type of fraud to which they are subjected. The OPLS-
DA technique should be prioritized because of the higher percentages of
3.5. Commercial samples analyses
specificity obtained.
The main advantage of PLS-DA and OPLS-DA is the possibility of
In order to evaluate the robustness and practical application of the
identifying the variables responsible for class identification. The regression
models developed samples of black pepper and cumin commercial
coefficients of the OPLS and PLS models were inspected and it was possible
powder were collected, processed and the results are presented in
to note the relevant variables to determine the presence of the adulterant in
Table 5.

10
A.B.S.d. Lima, et al. Food Control 107 (2020) 106802

The first observation that deserves attention is the high percentage 2136/sssaj2001.652480x.
of adulterated samples 62% and 79% for black pepper and cumin, re- Chen, H., Tan, C., & Lin, Z. (2018). Quantitative determination of wool in textile by near-
infrared spectroscopy and multivariate models. Spectrochimica Acta Part A: Molecular
spectively. While the pepper samples showed a higher percentage of and Biomolecular Spectroscopy, 201, 229–235. https://doi.org/10.1016/j.saa.2018.05.
adulteration by the addition compounds similar to cassava starch, the 010.
cumin samples were adulterated with corn flour analogs. In both spices, Ellis, D. I., Brewster, V. L., Dunn, W. B., Allwood, J. W., Golovanov, A. P., & Goodacre, R.
(2012). Fingerprinting food: Current technologies for the detection of food adul-
a higher level of adulteration was observed when corn flour or similar teration and contamination. Chemical Society Reviews, 41(17), 5706–5727. https://
were used in some samples, more than half of the spice consisted of an doi.org/10.1039/c2cs35138b.
adulterant. Possibly this fraudulent action can be really more employed FDA (2015). FDA consumer advice on products containing ground cumin with undeclared
peanuts. from http://wayback.archive-it.org/7993/20171114232613/https://www.
because of the facility that the corn presents in masking the visual and fda.gov/Food/RecallsOutbreaksEmergencies/SafetyAlertsAdvisories/ucm434274.
analytical characteristics of the fraud, being of difficult detection. These htm, Accessed date: 23 February 2019.
results are alarming and demonstrate the ease of performing this type of Fidelis, M., Santos, J. S., Coelho, A. L. K., Rodionova, O. Y., Pomerantsev, A., & Granato,
D. (2017). Authentication of juices from antioxidant and chemical perspectives : A
fraud in powder samples, reinforcing the relevance of the development
feasibility quality control study using chemometrics. Food Control, 73, 796–805.
of studies on the subject. The fraud rate on samples of pepper and https://doi.org/10.1016/j.foodcont.2016.09.043.
cumin powder is much higher than the authors had imagined at first. Galvin-King, P., Haughey, S. A., & Elliott, C. T. (2018). Herb and spice fraud; the drivers,
Therefore, the methods evaluated in the present study are good challenges and detection. Food Control, 88, 85–97. https://doi.org/10.1016/j.
foodcont.2017.12.031.
alternatives to predict purity or percentage of fraud with materials si- Garber, E. A. E., Parker, C. H., Handy, S. M., Cho, C. Y., Panda, R., Samadpour, M., et al.
milar to cassava starch and corn flour in new black pepper and cumin (2016). Presence of undeclared food allergens in cumin: The need for multiplex
samples, guaranteeing efficient and effective control of the quality of methods. Journal of Agricultural and Food Chemistry, 64(5), 1202–1211. https://doi.
org/10.1021/acs.jafc.5b05497.
these samples. Garcia, H., & Filzmoser, P. (2017). Multivariate Statistical Analysis using the R package
chemometrics. Austria: Vienna University of Technology. Retrieved from http://ftp.
4. Conclusion pregi.net/pub/R/web/packages/chemometrics/vignettes/chemometrics-vignette.
pdf.
Genisheva, Z., Quintelas, C., Mesquita, D. P., Ferreira, E. C., Oliveira, J. M., & Amaral, A.
This study indicates the feasibility and efficiency of the use of NIR L. (2018). New PLS analysis approach to wine volatile compounds characterization
spectroscopy associated to multivariate techniques to indicate the by near infrared spectroscopy (NIR). Food Chemistry, 246, 172–178. November 2017
https://doi.org/10.1016/j.foodchem.2017.11.015.
adulteration of black pepper and cumin by the addition of simple Gondaliya, S. B., Khatrani, T. J., Soni, K. K., & Baravalia, Y. K. (2018). Consequence on
amyloid (cassava starch) and complex (corn flour) sources. It shows long term storage on phytochemical attributes of cumin (Cuminum cyminum, L.)
that near-infrared spectroscopy combined with multivariate analysis from districts of north Gujarat, India. Industrial Crops and Products, 111(September
2017), 908–913. https://doi.org/10.1016/j.indcrop.2017.10.054.
can be a quick, simple and effective method of identifying and quan-
Hongyu, K., Sandanielo, V. L. M., & Oliveira Junior, G. J. (2015). Principal component
tifying spice fraud content and avoiding the use of expensive, time- analysis: Theory, interpretations and applications. Engineering and Science, 5(1),
consuming and complex traditional analytical techniques for this type 83–90. https://doi.org/10.18607/ES20165053.
of detection. Hübert, T., Tiebe, C., & Banach, U. (2016). Electronic noses for the quality control of spices.
Electronic noses and tongues in food science. Elsevier Inchttps://doi.org/10.1016/B978-
The models developed in this work will be kept “active” and up- 0-12-800243-8.00012-3.
dated to ensure their robustness for future analysis needs as new har- Jain, K. R., Modi, C. K., & Pithadiya, K. J. (2009). Non-destructive quality evaluation in spice
vests of the studied spices become available. industry with specific reference to cuminum cyminum L (cumin) seeds. (July).
John, P. (2012). Adulterated cumin seeds seized from 3 Unjha factories. from https://
timesofindia.indiatimes.com/city/ahmedabad/Adulterated-cumin-seeds-seized-
Declarations of interest from-3-Unjha-factories/articleshow/16871715.cms, Accessed date: 23 February
2019.
Kennard, R. W., & Stone, L. A. (1969). Computer aided design of experiments.
None. Technometrics, 11(1), 137–148. https://doi.org/10.1080/00401706.1969.10490666.
Lakshmi, V. (2012). Food adulteration. LABS, R. V., Guntur, Pradesh, A., & India
Acknowledgments International Journal of Science Inventions Today, 1(2), 106–113.
Lee, S., Lohumi, S., Cho, B.-K., Kim, M. S., & Lee, S.-H. (2014). Development of non-
destructive detection method for adulterated powder products using Raman spec-
The authors would like to thank the Coordenação de troscopy and partial least squares regression. Journal of the Korean Society for
Aperfeiçoamento Pessoal de Nível Superior (CAPES) and the Fundação Nondestructive Testing, 34(4), 283–289. https://doi.org/10.7779/jksnt.2014.34.4.
283.
de Amparo e Pesquisa do Estado da Bahia (FAPESB) for the granting of
Lee, S., Lohumi, S., Lim, H. S., Gotoh, T., Cho, B. K., Kim, M. S., et al. (2015).
Masters and Doctorate scholarships. Development of a detection method for adulterated onion powder using Raman
This research did not receive any specific grant from funding spectroscopy. Journal of the Faculty of Agriculture, 60(1), 151–156 Kyushu University.
agencies in the public, commercial, or not-for-profit sectors. Lohumi, S., Lee, S., & Cho, B. K. (2015). Optimal variable selection for Fourier transform
infrared spectroscopic analysis of starch-adulterated garlic powder. Sensors and
Actuators B: Chemical, 216, 622–628. https://doi.org/10.1016/j.snb.2015.04.060.
References Lohumi, S., Lee, S., Lee, W. H., Kim, M. S., Mo, C., Bae, H., et al. (2014). Detection of
starch adulteration in onion powder by FT-NIR and FT-IR spectroscopy. Journal of
Agricultural and Food Chemistry, 62(38), 9246–9251. https://doi.org/10.1021/
Bhat, U. R., & Tharanathan, R. N. (1983). Physicochemical properties of black pepper jf500574m.
(Piper nigrum) starch. Starch, 35(6), 189–192. https://doi.org/10.1002/star. Magwaza, L. S., Messo, S. I. N., Laurie, S. M., Laing, M. D., & Shimelis, H. (2016).
19830350603. Development of NIRS models for rapid quantification of protein content in sweet-
Black, C., Haughey, S. A., Chevallier, O. P., Galvin-King, P., & Elliott, C. T. (2016). A potato [Ipomoea batatas (L.) LAM.]. Lebensmittel-Wissenschaft und -Technologie- Food
comprehensive strategy to detect the fraudulent adulteration of herbs: The oregano Science and Technology, 72, 63–70. https://doi.org/10.1016/j.lwt.2016.04.032.
approach. Food Chemistry, 210, 551–557. https://doi.org/10.1016/j.foodchem.2016. Ma, X., Mao, W., Zhou, P., Li, P., & Li, H. (2015). Distinguishing foeniculum vulgare fruit
05.004. from two adulterants by combination of microscopy and GC – MS analysis. Microscopy
Bylesjo, M., Rantalainen, M., Cloarec, O., Nicholson, J. K., Holmes, E., & Trygg, J. (2006). Research and Technique, 78(May), 633–641. https://doi.org/10.1002/jemt.22523.
OPLS discriminant analysis : Combining the strengths of PLS-DA and SIMCA classi- McGrath, T. F., Haughey, S. A., Patterson, J., Fauhl-Hassek, C., Donarski, J., Alewijn, M.,
fication y. Journal of Chemometrics, 20(February), 341–351. https://doi.org/10.1002/ et al. (2018). What are the scientific challenges in moving from targeted to non-
cem. targeted methods for food fraud testing and how can they be addressed? – spectro-
Cevoli, C., Gianotti, A., Troncoso, R., & Fabbri, A. (2015). Quality evaluation by physical scopy case study. Trends in Food Science & Technology, 76, 38–55. https://doi.org/10.
tests of a traditional Italian flat bread Piadina during storage and shelf-life im- 1016/j.tifs.2018.04.001.
provement with sourdough and enzymes. European Food Research and Technology, Mehrotra, R. (2006). Infrared spectroscopy, gas chromatography/infrared in food ana-
240(6), 1081–1089. https://doi.org/10.1007/s00217-015-2429-7. lysis. Encyclopedia of Analytical Chemistry, 1–18https://doi.org/10.1002/
Chakravartula, S. S. N., Cevoli, C., Balestra, F., Fabbria, A., & Rosa, M. D. (2019). 9780470027318.a1013.
Evaluation of drying of edible coating on bread using NIR spectroscopy. Journal of Moyer, D. C., DeVries, J. W., & Spink, J. (2017). The economics of a food fraud incident –
Food Engineering, 240, 29–37. https://doi.org/10.1016/j.jfoodeng.2018.07.009. case studies and examples including Melamine in Wheat Gluten. Food Control, 71,
Chang, C.-W., Laird, D. A., Mausbach, M. J., & Hurburgh, C. R., Jr. (2001). Analyses of 358–364. https://doi.org/10.1016/j.foodcont.2016.07.015.
soil properties. Soil Science Society of America Journal, 480–490. https://doi.org/10. Ncama, K., Magwaza, L, S., Mditshwa, A., & Z. Tesfay, S. (2019). Application of visible to

11
A.B.S.d. Lima, et al. Food Control 107 (2020) 106802

near-infrared spectroscopy for non-destructive assessment of quality parameters of in Analytical Chemistry, 78, 17–22. https://doi.org/10.1016/j.trac.2016.01.010.
fruit. Infrared spectroscopy - principles, advances, and applications (pp. 1–17). . https:// Sampaio, P., Soares, A., Castanho, A., Almeida, A. S., Oliveira, J., & Brites, C. (2017).
doi.org/10.5772/intechopen.80069. Dataset of Near-infrared spectroscopy measurement for amylose determination using
Oladunmoye, O. O., Aworh, O. C., Maziya-Dixon, B., Erukainure, O. L., & Elemo, G. N. PLS algorithms. Data in Brief, 15, 389–396. https://doi.org/10.1016/j.dib.2017.09.
(2014). Chemical and functional properties of cassava starch, durum wheat semolina 077.
flour, and their blends. Food Sciences and Nutrition, 2(2), 132–138. https://doi.org/ Schweiggert, U., Carle, R., & Schieber, A. (2007). Conventional and alternative processes
10.1162/089976602753284446. for spice production - a review. Trends in Food Science & Technology, 18(5), 260–268.
Oliveri, P., Casale, M., Casolino, M. C., Baldo, M. A., & Grifi, F. N. (2011). Comparison https://doi.org/10.1016/j.tifs.2007.01.005.
between classical and innovative class-modelling techniques for the characterisation Sicherer, S. H., & Sampson, H. A. (2010). Food allergy. The Journal of Allergy and Clinical
of a PDO olive oil. Analitical and Bioanalitical Chemistry, 399, 2105–2113. https://doi. Immunology, 125(2), 355–366. https://doi.org/10.1007/978-981-10-5541-6_28.
org/10.1007/s00216-010-4377-1. Silvis, I. C. J., van Ruth, S. M., van der Fels-Klerx, H. J., & Luning, P. A. (2017).
Oliveri, P., & Downey, G. (2012). Multivariate class modeling for the verification of food- Assessment of food fraud vulnerability in the spices chain: An explorative study. Food
authenticity claims. Trends in Analytical Chemistry, 35, 74–86. https://doi.org/10. Control, 81, 80–87. https://doi.org/10.1016/j.foodcont.2017.05.019.
1016/j.trac.2012.02.005. Tabachnick, B. G., & Fidell, L. S. (2006). Using multivariate statistics (5th ed.). MA: Inc.
Pasquini, C. (2018). Near infrared spectroscopy: A mature analytical technique with new Needham Heights.
perspectives – a review. Analytica Chimica Acta, 1026, 8–36. https://doi.org/10. Vadivel, V., Ravichandran, N., Rajalakshmi, P., Brindha, P., Gopal, A., & Kumaravelu, C.
1016/j.aca.2018.04.004. (2018). Microscopic, phytochemical, HPTLC, GC–MS and NIRS methods to differ-
Qamar, S., Aslam, M., Huyop, F., & Javed, M. A. (2017). Comparative study for the de- entiate herbal adulterants: Pepper and papaya seeds. Journal of Herbal Medicine, 11,
termination of nutritional composition in commercial and noncommercial maize 36–45. https://doi.org/10.1016/j.hermed.2018.01.004.
flours. Pakistan Journal of Botany, 49(2), 519–523. Wilde, A. S., Haughey, S. A., Galvin-King, P., & Elliott, C. T. (2019). The feasibility of
Ramos-Diaz, J. M., Rinnan, Å., & Jouppila, K. (2019). Application of NIR imaging to the applying NIR and FT-IR fingerprinting to detect adulteration in black pepper. Food
study of expanded snacks containing amaranth, quinoa and kañiwa. Lebensmittel- Control, 100, 1–7. https://doi.org/10.1016/j.foodcont.2018.12.039.
Wissenschaft & Technologie, 102(November), 8–14. 2018 https://doi.org/10.1016/j. Williams, P. C., & Sobering, D. C. (1993). Comparison of commercial near infrared
lwt.2018.12.029. transmittance and reflectance instruments for analysis of whole grains and seeds.
Rodionova, O. Y., Titova, A. V., & Pomerantsev, A. L. (2016). Trends in Analytical Journal of Near Infrared Spectroscopy, 1(1), 25–32. https://doi.org/10.1255/jnirs.3.
Chemistry Discriminant analysis is an inappropriate method of authentication. Trends

12