Food Chemistry 194 (2016) 441446

Contents lists available at ScienceDirect

Food Chemistry
journal homepage: www.elsevier.com/locate/foodchem

Analysis of spreadable cheese by Raman spectroscopy and chemometric

tools
Kamila de S Oliveira a, Layce de Souza Callegaro a, Rodrigo Stephani b, Mariana Ramos Almeida a,
Luiz Fernando Cappa de Oliveira a,
a
Ncleo de Espectroscopia e Estrutura Molecular (NEEM), Departamento de Qumica, Universidade Federal de Juiz de Fora, 36036-330 Juiz de Fora, MG, Brazil
b
Tate & Lyle Gemacom Tech, 36092-050 Juiz de Fora, MG, Brazil

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

Article history: In this work, FT-Raman spectroscopy was explored to evaluate spreadable cheese samples. A partial least
Received 15 June 2015 squares discriminant analysis was employed to identify the spreadable cheese samples containing starch.
Received in revised form 7 August 2015 To build the models, two types of samples were used: commercial samples and samples manufactured in
Accepted 11 August 2015
local industries. The method of supervised classification PLS-DA was employed to classify the samples as
Available online 12 August 2015
adulterated or without starch. Multivariate regression was performed using the partial least squares
method to quantify the starch in the spreadable cheese. The limit of detection obtained for the model
Keywords:
was 0.34% (w/w) and the limit of quantification was 1.14% (w/w). The reliability of the models was eval-
Spreadable cheese
Raman spectroscopy
uated by determining the confidence interval, which was calculated using the bootstrap re-sampling
Quality control technique. The results show that the classification models can be used to complement classical analysis
Starch and as screening methods.
PLS-DA 2015 Elsevier Ltd. All rights reserved.

1. Introduction
Milk and dairy products are important nutrient sources and are
considered primary sources of biological calcium, which is needed
for bone mass formation. Dairy products are constant targets of
economic adulteration due to their high demand and seasonal
price variations. In these products, ingredients that are not speci-
fied on the label or are not allowed are frequently found.
Therefore, it is important to ensure fair competition between com-
panies and protect consumers against fraud (Karoui & De
Baerdemaeker, 2007).
An important Brazilian dairy product is spreadable cheese,
which has seen increasing consumption since the 1990s. Today,
it is the third best selling dairy product (ABIQ, 2014). Spreadable
cheese is a Brazilian processed cheese, originating as a homemade
means of utilizing spontaneously clotted milk due to the action of
milks natural microbiota. It is made from raw or pasteurized
skimmed milk, with or without the addition of lactic cultures
(Van Dender, 2014).
Starch has been extensively studied for use in processed
cheeses because its physical properties, including its rheology
and microstructure, are influenced by interactions between the
Corresponding author.
E-mail address: luiz.oliveira@ufjf.edu.br (L.F.C. de Oliveira).
starch and milk proteins. However, the main benefit of adding
starch to dairy products is in the cost reduction (Van Dender,
2014). The addition of starch can lead to lower pricing and result
in unfair competition, hampering legitimate business. When the
product includes added starch, it must be indicated on the main
panel of the label.
Today, Normative Instruction 68/2006 of the Ministry of
Agriculture, Livestock and Food Supply regulates the official meth-
ods used for the physicochemical analysis of milk products. This
includes a qualitative analysis of starch using the Lugol test and
a quantitative analysis based on the LaneEynon method. As part
of this method, a clarification and filtration of the dairy product
is first performed to separate the simple and complex sugars.
Then, an acid digestion of the solid waste is conducted, a different
clarification and filtration process is used. Finally, this second fil-
trate is subjected to titration with the Fehling reagent (Wehr &
Frank, 2004).
However, the quantitative method can cause some problems. In
addition to demanding the use of many reagents, it requires skilled
labour and generates waste. Additionally, there is evidence of a loss
of starch during the first filtration (Vallado, 2012). Thus, it is dif-
ficult to obtain reliable results to identify the presence of starch in
spreadable cheese.
In recent decades, spectroscopic methods coupled with
chemometric tools have been developed to evaluate the quality

http://dx.doi.org/10.1016/j.foodchem.2015.08.039
0308-8146/ 2015 Elsevier Ltd. All rights reserved.
442 K. de S Oliveira et al. / Food Chemistry 194 (2016) 441446
of milk products as an alternative to the reference procedures.
Borin, Ferro, Mello, Maretto, and Poppi (2006) proposed a
method for starch, sucrose and whey quantification, which are
common adulterants in milk powder using near infrared
spectroscopy and artificial intelligence techniques. Karoui and
De Baerdemaeker (2007) discussed the potential of destructive
and non-destructive techniques coupled with chemometric tools
to determine the quality and authenticity of dairy products.
Almeida, Oliveira, Stephani, and De Oliveira (2011, 2012) used
Raman spectroscopy and chemometric tools to assess the quality
and authenticity of milk powder of whey and starch as
adulterants.
Due to the lack of efficient methods for spreadable cheese anal-
ysis, Raman spectroscopy is a promising technique due to sample
preparation methods and the ability to obtain chemical informa-
tion. Nevertheless, the employment of Raman spectroscopy in ana-
lytical chemistry requires the use of chemometric tools; a
chemometric analysis can be used to create automatic classifica-
tion models, as well as to validate the measured data and provide
more reliable information.
Although many publications describe several applications of
Raman spectroscopy and chemometric analyses, such approach
are not yet employed in standard laboratories and industries.
Usually they prefer NIR spectroscopy because is cheaper and sim-
ple. However, Raman spectroscopy provides the higher informa-
tion content, since the spectrum contains information of each
one of the chemical components. Other reasons for the low practi-
cal applicability include difficulties in implementing such analyses
and a lack of analytical validation for the developed models. Every
time an analysis method is developed, the results must be vali-
dated. In light of this, this investigation calculates the confidence
intervals for the classification models and multivariate calibration
using the bootstrap re-sampling technique (Efron & Tibshirani,
1994).
In this work, we propose a qualitative and quantitative
analysis of starch in commercial spreadable cheese products
by employing Raman spectroscopy and chemometric tools.
Classification models employing the Partial Least Squares
Discriminant Analysis (PLS-DA) were used to classify adulterated
and unadulterated spreadable cheese samples, and the Partial
Least Squares (PLS) method were applied to construct a quantita-
tive model.
2. Materials and methods
2.1. Samples
To construction of the multivariate models, 11 spreadable
cheese samples without reducing fats were prepared by
Gemacom Tech, which is headquartered in the city of Juiz de
Fora, MG. Among these, one is a standard sample without the
addition of starch and the other ten were divided into two groups
according the added starch. Five samples were adulterated with
starch containing 12% of amylose and 88% of amylopectin,
referred to as starch 2560, and five were adulterated with starch
containing <5% amylose and a balance of amylopectin, referred to
as starch 4051. In both groups, the adulterated samples were
prepared with a variation from 2% to 10% (w/w) and divided into
five concentration levels, as shown in Table S1 (Supplementary
Material).
A total of 27 samples of different brands of commercial spread-
able cheese were also analysed, including 15 whole samples and 12
skimmed samples. These were purchased at a local market in the
city of Juiz de Fora.
2.2. FT-Raman measurements
The Raman spectra of the spreadable cheese samples and starch
were collected on a RFS 100 FT-Raman Bruker spectrometer
equipped with a Ge detector using liquid nitrogen as the coolant
and a Nd:YAG laser emitting at 1064 nm. The laser light, with a
power of 150 mW, was introduced and focused on the sample,
and the scattered radiation was collected at 180. For each spec-
trum, an average of 256 scans was performed at a resolution of
4 cm 1 over a range from 3500 to 50 cm 1. The OPUS 6.0 (Bruker
Optik, Ettlingen, Germany) software was used for Raman data
acquisition. For all of the FT-Raman spectra obtained in this work,
the samples did not undergo any previous preparation. The Raman
spectra were obtained in triplicate for the spreadable cheese.
2.3. Chemometric analysis
To perform the multivariate analysis, the Matlab 7.10.0 (R2010a)
software was used. For all of the models, the spectral range used
was 3500400 cm 1. The spectra (matrix X) were processed
employing the first derivative with the SavitzkyGolay smoothing
algorithm (Savitzky & Golay, 1964) using a 9-point window and a
second-order polynomial function. In addition, normalization by
length and the mean centring of data were also applied.
After pre-processing, PCA was performed for a group composed
of all of the samples (93 samples) including those that prepared
in the laboratory (Gemacom Tech) and those that purchased at the
local market. The choice of the number of principal components
(PCs) was determined through the eigen-values graph (which rep-
resents the importance of each principal component).
For the construction of classification models using PLS-DA, the
samples were divided into two subsets: calibration and validation.
The calibration subset included 61 samples and the validation sub-
set included 32 samples. The subsets contain adulterated and
unadulterated samples. A vector Y, containing only the numbers
0 and 1 was created to indicate the class each sample belonged
to: 0 for unadulterated and 1 for adulterated spreadable cheese.
The criterion for the lowest prediction error in cross-validation
was used to select the number of latent variables for the PLS-DA
model. The threshold value for the class separation was calculated
using Bayes Theorem. This value was selected as the point where
the two estimated distributions cross, which is the y predicted
value at which the number of false responses is minimized for
the new predictions. After building the classification rules with
the calibration set, the model was evaluated with the validation
set and the performance was estimated in terms of the correct clas-
sification of the samples, sensibility, specificity, efficiency and
Matthews correlation coefficient (Almeida, Fidelis, Barata, &
Poppi, 2013). The parameters mentioned give an overview of the
behaviour of the model. To assess the uncertainties of each sample
the confidence interval was estimated using the residual bootstrap
technique (Efron & Tibshirani, 1994), according to Almeida et al.
(2013).
After identifying the adulterated samples, the starch content in
the spreadable cheese was calculated using the Partial Least
Squares (PLS) Regression. The calibration and validation subsets
from the classification models were also used in this case, deleting
only commercial samples containing starch because the proportion
of added starch was not provided on the label. The matrices con-
taining the data provided by the Raman spectra and the vector Y
with the concentrations of added adulterant to each sample were
employed to build the regression model. The number of latent vari-
ables was chosen by cross-validation (RMSECV value). The perfor-
mance of the PLS model was evaluated with the validation samples
using parameters such as accuracy (RMSEP), limit of detection
(LOD), limit of quantification (LOQ), linearity, model fit and
 K. de S Oliveira et al. / Food Chemistry 194 (2016) 441446
uncertainties. The detection limit and the limit of quantification
were estimated by calculating the net analyte signal (NAS)
(Lorber, 1986). The uncertainties in the prediction of the PLS model
were calculated using the residual bootstrap.
3. Results and discussion
3.1. FT-Raman spectra of spreadable cheese samples
Two of the FT-Raman spectra obtained for the samples of whole
and skimmed spreadable cheeses are shown in Fig. 1. The main dif-
ference between the two types of spreadable cheese is the partial
or total lipid presence, which is the nutrient that determines its
classification as either whole or skimmed. This difference is shown
in the Raman spectrum by bands at 2886 and 2852 cm 1, which
are related to the vibrational modes of lipids. In the whole spread-
able cheese spectrum, these appear with a higher intensity with
respect to the band at 2929 cm 1, indicating the major quantity
of this nutrient in comparison to the skimmed samples (Li-Chan,
1996).
The 1800400 cm 1 region features a large number of bands
and is rich in structural information. The band at 1745 cm 1 is
assigned to C@O stretching (Zhou et al., 2006) and is characteristic
of a lipid vibrational mode present in a lower intensity in the
Raman spectrum of skimmed spreadable cheese. The band at
1652 cm 1 is characteristic of the amide I mode and is related to
the C@O stretching of protein molecules. This region also con-
tributes to the C@C stretching of unsaturated fatty acids present
in milk (Moros, Garrigues, & De la Guardia, 2007; Zhou et al.,
2006). An intense band at 1440 cm 1 is observed and assigned to
the deformation modes of the CH2 group. This band is characteris-
tic of lipids, carbohydrates and aliphatic amino acids present in
milk protein (Li-Chan, 1996; Zhou et al., 2006). Another band
related to the presence of lipids appears at 1299 cm 1 and is
referred to as the CH2 twisting mode (Herrero, 2008).
In the spectral range from 1200 to 800 cm 1, characteristic
vibrational modes appear related to carbohydrates. The vibrational
modes present are assigned to CO stretching, the CC and COH
deformation modes (11201064 cm 1) and COC deformation
(950870 cm 1), which is related to the vibrational mode of the
b-1-4 glycosidic bond (Li-Chan, 1996; Mahdad-Benzerdjeb, Taleb-
Mokhtari, & Sekkal-Rahal, 2007).
Fig. 1. FT-Raman spectra of: (a) skimmed and (b) whole spreadable cheese. The
experimental conditions are described in Section 2.
443
Fig. 2. FT-Raman spectra of: (a) standard spreadable cheese, (b) 2% (w/w) starch, (c)
6% (w/w) starch, (d) 10% (w/w) starch and (e) pure starch. The experimental
conditions are described in Section 2.
Fig. 2 shows the spectral region between 2000 and 400 cm 1 for
pure standard spreadable cheese, with the addition of 2%, 6% and
10% (w/w) of starch and the starch itself, used as adulterant. As
the carbohydrate concentration increases, there is a slight modifi-
cation in the band at 477 cm 1 assigned to the [d(CCC) + s(CO)]
mode (Almeida et al., 2010), which is described in the literature
as a marker of the presence of starch (Baranska, Schulz, Baranski,
Nothnagel, & Christensen, 2005; De Veij, Vandenabeele, De Beer,
Remonc, & Moens, 2009).
The visual analysis of the Raman spectra of the spreadable
cheese allows the identification of the presence of starch from a
concentration of 6% (w/w). Below this concentration limit, chemo-
metric tools are required.
3.2. Exploratory analysis using PCA of the FT-Raman spectra of
spreadable cheese samples
An option for spectral pre-processing is the normalization
method for a vector of unit length. This was applied to the
first derivative using SavitzkyGolay smoothing with a 9-point
window and a second-order polynomial for smoothing and
mean-centring the data. This pre-processing provides better
classification analysis performance and yields the best interpreta-
tion of the spectra.
PCA was employed to examine the formation of the sample
groups with similar characteristics. In the score analysis of the first
three principal components, which represent 71.58% of the original
information, it is possible to observe a split between the skimmed
and whole spreadable cheese samples, with or without starch
(Fig. S1, Supplementary Material). Such a division is promoted by
the first principal component and can be confirmed by examining
the loading graph (Fig. S2, Supplementary Material). It is notable
that the most relevant regions of the spectrum for the division
are related to the vibrational modes of the lipids, including 2886,
2852 and 1440 cm 1, whose vibrational assignments are discussed
in Section 3.1. Moreover, the distribution of the samples is accord-
ing to the starch presence was determined by the third principal
component. The loading plot (Fig. S3, Supplementary Material)
shows that one of the most relevant spectral regions for the distri-
bution is at 477 cm 1 and is used in this study as a marker band for
starch. Additionally, based on the exploratory analysis, we con-
clude that despite having used two different types of starch in
the manufacture of the spreadable cheese, there are no spectral dif-
ferences in the samples because all of the adulterated samples pre-
pared in the laboratory were allocated into the same grouping.
444 K. de S Oliveira et al. / Food Chemistry 194 (2016) 441446
The PCA model did not provide an adequate discrimination
between samples with or without additional starch, mainly at
low concentrations. Thus, it requires the employment of a super-
vised classification method, with an initial knowledge about the
classes to be modelled.
3.3. Discriminant analysis
The PLS-DA supervised classification method was employed to
obtain better discrimination of the adulterated and non-
adulterated samples. For the construction of the models, the sam-
ples were divided as described in Section 2. Three latent variables
(LV) were selected for the development of the classification model,
representing 78% of the variance explained in the X matrix (the
spectra data set) and 90% of the variance in the Y vector.
The results of the classification models for the calibration and
validation set are shown in Fig. 3a and b, respectively. The calcu-
lated threshold is shown as the horizontal line in Fig. 3, with a
value of 0.59. Above this value, spreadable cheese is classified as
adulterated with starch. Below this value indicates no starch addi-
tion. The threshold choice is very important, as it determines the
model quality. If we increase the threshold value, the probability
Fig. 3. Results of the: (a) calibration and (b) validation set of the PLS-DA model,
showing spreadable cheese with starch addition (above the dashed line) and
spreadable cheese without starch (below the dashed line) with confidence intervals
calculated using the bootstrap re-sampling method.
of obtaining a false positive is reduced, but any false negative
results are augmented. In this work, it is important to avoid false
negatives, as that would classify spreadable cheese with starch as
without any addition.
All samples shown in Fig. 3, either adulterated or not, were clas-
sified correctly. This result shows excellent performance for the
studied samples with 100% accuracy, sensitivity and specificity
and a Matthews correlation coefficient of 1. Thus, the PLS-DA
models is useful for analysing new samples of spreadable cheese
obtained using Raman spectroscopy data, predicting if they contain
starch or not.
The parameters evaluated give a general vision of the behaviour
of the model. To obtain information about each individual sample,
the uncertainty in the classification for each sample was computed
using the residual bootstrap. The results are the error bars shown
in Fig. 3. The results show that the samples in the calibration and
validation set were correctly classified with 95% confidence.
The classification model was based on the profile of the Raman
spectra of the spreadable cheese and the analysis of the loading of
the latent variables allow identification of the Raman bands that
contribute to the separation between the samples with and with-
out starch addition. The loading plots for the three latent variables
are presented in Fig. 4 and show that the regions around 2880,
1750, 1650, 1440 and 477 cm 1 are the most important to the clas-
sification model. The first two regions (2880 and 1750 cm 1) are
related to the vibrational modes of lipids, whose vibrational
assignments are discussed in Section 3.1, whereas the band at
477 cm 1 is due to the presence of starch and is promoted by the
second latent variable.
3.4. Quantification models
After identifying the presence of starch in spreadable cheese
using the PLS-DA model, a quantification model based on the
Partial Least Squares (PLS) was constructed for determination of
the starch content in the adulterated samples. For this, the samples
were divided into calibration and validation sets as described in
Section 2. Three latent variables (LVs) were employed, which
explained 73% of the total variance in the X matrix and 98% of
the total variance in the Y vector.
Fig. 4. Loading plot for the LV 1, LV 2 and LV 3 versus wavenumber (variables) for
the PLS-DA model.
 K. de S Oliveira et al. / Food Chemistry 194 (2016) 441446
Fig. 5. Predicted values from the PLS model versus the reference values of % starch
(% w/w) added to spreadable cheese: (a) calibration set and (b) validation set. The
uncertainties in terms of % starch were calculated using the bootstrap re-sampling
method.
Fig. 5 show the correlation between the references values
(amount of starch added in the laboratory) and those predicted
by the PLS regression. The correlation between the reference values
and those estimated by the PLS model for the calibration samples
was 0.989 and 0.984 for the validation samples. These fits are in
accordance with the requirements from regulatory agencies in
Brazil.
The model was evaluated via the calculation of figures of merit,
with the results presented in Table 1. The accuracy was assessed by
the values of the root mean squared error for the calibration
(RMSEC) and validation (RMSEP), which were 0.58% (w/w) for
the RMSEC and 0.65% (w/w) for the RMSEP. These values show a
good estimated starch content when using the PLS model. The lin-
earity is assessed by residual models: the calibration samples
exhibited an absence of systematic trends in the distribution of
the residuals. Thus, the previously discussed correlation coefficient
can be used to evaluate the linearity of the model. For the estima-
tion of other figure of merits shown in Table 1, the NAS concept
was employed. The limit of detection obtained for the model was
0.34% (w/w), and the limit of quantification was 1.14% (w/w).
The proposed quantification model can detect starch values higher
445
Table 1
Results for the optimization and parameters estimated for calculating the primary
figures of merit for the PLS model.
Optimization
Number of calibration set 57
Number of validation set 29
Number of latent variables 3
Figures of merit
Accuracy RMSECV 0.66
RMSEC 0.58
RMSEP 0.65
Linearity Slope 0.98
Intercept 0.08
Correlation coefficient 0.98
Limit of detection (w/w) 0.34
Limit of quantification (w/w) 1.14
than 0.34% (w/w) and quantify starch values above 1.14% (w/w) in
adulterated spreadable cheese.
The 95% confidence intervals were calculated using bootstrap
re-sampling. It was observed that the uncertainties are greater
when smaller amounts of starch are present in spreadable cheese.
For some samples without the addition of starch, the values
obtained for the PLS regression were negative. This problem was
also observed and reported in another study involving PLS regres-
sion and molecular spectroscopy data (Borin et al., 2006). However,
this is not an obstacle in this work because the quantification is
used to quantify samples exhibiting the presence of starch after
identification using the classification model.
4. Conclusions
In this work, we proposed an analytical methodology to charac-
terize spreadable cheeses (whole and skimmed) and identify added
starch in spreadable cheese using Raman spectroscopy. The charac-
terization of whole and skimmed spreadable cheeses is very sim-
ple; the differentiation is based on the presence of a band at
1745 cm 1 that is characteristic of lipids, which are absent in
skimmed samples. The Raman spectra analysis of spreadable
cheeses allows the visual identification of the presence of starch
at a minimum concentration of 6% (w/w). The detection of smaller
quantities requires chemometric tools.
The classification model exhibits excellent performance with all
of the samples being correctly classified. The uncertainty calcula-
tion during sample classification was performed using the boot-
strap re-sampling procedure and allowed the construction of
more reliable classification model, decreasing the chances of mis-
classification. After identifying starch in spreadable cheeses, a
quantification model based on the Partial Least Squares (PLS) was
constructed for determining the starch content in the samples.
The limit of detection obtained in this model was 0.34% (w/w),
and the limit of quantification was 1.14% (w/w). The 95% confi-
dence interval was also calculated using the bootstrap re-
sampling method.
The developed methodology may be used as a screening test in
the quality control of spreadable cheese, as well as other types of
soft cheese; the procedure also presents the possibility of on-line
measurements, employing portable Raman instruments.
Acknowledgements
The authors would like to thank CAPES, FAPEMIG and CNPq
(Brazilian agencies) for their financial support, as well as Tate &
Lyle Gemacom Tech for public-private partnership in research.
This work is a collaborative research project with members of
Rede Mineira de Qumica (RQ-MG) supported by FAPEMIG
(Project: CEX RED-0010-14).
446 K. de S Oliveira et al. / Food Chemistry 194 (2016) 441446
Appendix A. Supplementary data
Supplementary data associated with this article can be found, in
the online version, at http://dx.doi.org/10.1016/j.foodchem.2015.
08.039.
References
ABIQ (Brazilian Association of Cheese Industries) (2014). URL: <http://www.
abiq.com.br/default.asp> Accessed 12.04.15.
Almeida, M. R., Alves, R. S., Nascimbem, L. B. L. R., Stephani, R., Poppi, R. J., & De
Oliveira, L. F. C. (2010). Determination of amylose content in starch using
Raman spectroscopy and multivariate calibration analysis. Analytical and
Bioanalytical Chemistry, 397, 26932701.
Almeida, M. R., Fidelis, C. H. V., Barata, L. E. S., & Poppi, R. J. (2013). Classification of
Amazonian rosewood essential oil by Raman spectroscopy and PLS-DA with
reliability estimation. Talanta, 117, 305311.
Almeida, M. R., Oliveira, K. S., Stephani, R., & De Oliveira, L. F. C. (2011). Fourier-
transform Raman analysis of milk powder: A potential method for rapid quality
screening. Journal of Raman Spectroscopy, 42, 15481552.
Almeida, M. R., Oliveira, K. S., Stephani, R., & De Oliveira, L. F. C. (2012). Application
of FT-Raman spectroscopy and chemometric analysis for determination of
adulteration in milk powder. Analytical Letters, 45, 25892602.
Baranska, M., Schulz, H., Baranski, R., Nothnagel, T., & Christensen, L. P. (2005). In
situ simultaneous analysis of polyacetylenes, carotenoids and polysaccharides
in carrot roots. Journal of Agricultural and Food Chemistry, 53, 65656571.
Borin, A., Ferro, M. F., Mello, C., Maretto, D. A., & Poppi, R. J. (2006). Least-squares
support vector machines and near infrared spectroscopy for quantification of
common adulterants in powdered milk. Analytica Chimica Acta, 579, 2532.
De Veij, M., Vandenabeele, P., De Beer, T., Remonc, J. P., & Moens, L. (2009).
Reference database of Raman spectra of pharmaceutical excipients. Journal
Raman Spectroscopy, 40, 297307.
Efron, B., & Tibshirani, R. J. (1994). An Introduction to the Bootstrap. Boca Raton:
Chapman and Hall/CRC (pp. 456).
Herrero, A. M. (2008). Raman spectroscopy a promising technique for quality
assessment of meat and fish: A review. Food Chemistry, 107, 16421651.
Karoui, R., & De Baerdemaeker, J. (2007). A review of the analytical methods coupled
with chemometric tools for the determination of the quality and identity of
dairy products. Food Chemistry, 102, 621640.
Li-Chan, E. C. Y. (1996). The applications of Raman spectroscopy in food science.
Trends in Food Science & Technology, 7, 361370.
Lorber, A. (1986). Error propagation and figures of merit for quantification by
solving matrix equations. Analytical Chemistry, 58, 11671172.
Mahdad-Benzerdjeb, A., Taleb-Mokhtari, I. N., & Sekkal-Rahal, M. (2007). Normal
coordinates analyses of disaccharides constituted by D-glucose, D-galactose and
D-fructose units. Spectrochimica Acta Part A, 68, 284299.
Moros, J., Garrigues, S., & De la Guardia, M. (2007). Evaluation of nutritional
parameters in infant formulas and powdered milk by Raman spectroscopy.
Analytical Chimica Acta, 593, 3038.
Savitzky, A., & Golay, M. J. E. (1964). Smoothing and differentiation of data by
simplified least squares procedures. Analytical Chemical, 36, 16271639.
Vallado, S. A. (2012). Evaluation of the official method IN 68/2006 for the analysis of
starch in yogurt (Masters Dissertation). Federal University of Minas Gerais,
Faculty of Pharmacy, Graduate Program in Food Science.
Van Dender, A. G. F. (2014). Requeijo cremoso e outros queijos fundidos: tecnologia de
fabricao, controle do processo e aspectos de mercado (2nd ed.). So Paulo:
Setembro.
Wehr, H. M., & Frank, J. F. (2004). Standard methods for the examination of dairy
products (17th ed.). Washington, DC: American Public Health Association.
Zhou, Q., Sun, S. Q., Yu, L., Xu, C. H., Noda, I., & Zhang, X. R. (2006). Sequential
changes of main components in different types of milk powders using two-
dimensional infrared correlation analysis. Journal of Molecular Structure, 799,
7784.