Vibrational Spectroscopy 67 (2013) 8791

Contents lists available at SciVerse ScienceDirect

Vibrational Spectroscopy
journal homepage: www.elsevier.com/locate/vibspec

Raman spectroscopic quantication of calcium carbonate in spiked milk powder

samples
G.P.S. Smith a , K.C. Gordon a, , S.E. Holroyd b
a
b

Department of Chemistry, University of Otago, Dunedin, New Zealand

Fonterra Research and Development Centre, Palmerston North, New Zealand

a r t i c l e

i n f o

Article history:
Received 26 October 2012
Received in revised form 19 April 2013
Accepted 24 April 2013
Available online 13 May 2013
Keywords:
Raman spectroscopy
Milk powder
Calcium carbonate
Partial least squares

a b s t r a c t
Raman spectroscopy was used to determine the limits of detection and quantication (LOD and LOQ,
respectively) of calcite in spiked milk powder samples. Samples of milk powder spiked with 0.510%
(weight by weight; w/w) calcite were prepared and analysed using FT-Raman spectroscopy. Using the
spectra obtained from these samples, calibration models were prepared using peak intensities, analysis
of peak intensity ratios, based on band integrals and peak heights, and multivariate analysis. Analysis
methods using peak intensity, peak integral ratios and peak height ratios produced LOD values of 3.8, 3.3
and 1.4% (w/w) calcite and LOQ values of 13, 11 and 4.7% (w/w) calcite respectively. Multivariate analysis
produced a LOD of 1.0% (w/w) calcite and a LOQ of 3.4% (w/w) calcite.
2013 Elsevier B.V. All rights reserved.

1. Introduction
A large quantity of milk powder is consumed every year globally. Over 3.3 million metric tonnes are estimated to be consumed
in 2012; almost half of which is expected to be consumed in China
alone [1]. Milk powder typically contains approximately 25.2% protein, 38.1% carbohydrates, 26.2% fat, 7% ash and 3.5% water [2]. Fat,
carbohydrates and protein are important to human nutrition as fat
and carbohydrates are excellent sources of energy and protein provides energy as well as a source of essential amino acids [3]. Ash
consists of the oxides of minerals obtained after combustion. Milk
powder has several advantages over its liquid counterpart including: longer shelf-life and the ability to give a larger quantity of milk
per unit of mass when reconstituted with potable water. Calcite is
a common food additive used to increase the calcium content of
various foods, including milk powder, in order to promote healthy
bodily functions such as cellular metabolism and enzyme activation
[4,5]. Calcium carbonate has been identied in milk powder using
Raman spectroscopy in a previous study, which also reported the
efcacy of Raman spectroscopy for quantifying other milk powder
constituents, namely milk fat and protein [6]. The effect of calcium
carbonate on the heat stability of skim milk powder has also been
investigated [5].
Mid-infrared (MIR) and near infrared (NIR) have been used in
many previous studies in the analysis of milk powder constituents,

Corresponding author. Tel.: +64 3479 7599; fax: +64 3479 7906.
E-mail address: kgordon@chemistry.otago.ac.nz (K.C. Gordon).
0924-2031/$ see front matter 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.vibspec.2013.04.005

with much of the focus on constituent quantication [713].

Recently, several studies have been performed on dairy products using Raman spectroscopy. Composition of milk fat globules,
prediction of solid fat and triacylglycerol content, and the quantication of fat in milk have all been investigated recently using this
technique [1416]. An investigation of milk powder constituent
quantication has also been performed by McGoverin et al. [6].
Quantication of whey added to milk powder has also been investigated [17]. Use of Raman spectroscopy in quality control has
already been considered in the investigation of infant formula
laced with melamine due to several advantages. This technique has
compared with previously used techniques such as enzyme-linked
immunosorbent assays and high performance liquid chromatography [18]. Raman spectroscopy is quick and easy to use, requiring
little to no prior preparation of samples and is not hindered by
the presence of water in the samples like infrared spectroscopy
[18,19]. It should also be noted that use of univariate analysis
for quantication of sample constituents has been investigated
on many occasions [2022]. The method for producing univariate
regressions often varies between investigations, suggesting several
quantication methods could be employed in this study [2022].
In the work presented here, Raman spectroscopic data were
collected from milk powder samples which had been spiked with
calcium carbonate. These data were then processed using three univariate analysis methods and partial least squares (PLS) analysis
in order to determine whether calcium carbonate content can be
quantied. This involved construction of prediction models as well
as calculation of the limits of detection and quantication (LOD and
LOQ, respectively).

88

G.P.S. Smith et al. / Vibrational Spectroscopy 67 (2013) 8791

2. Materials and methods

2.1. Milk powder samples
Commercially available Pams Instant Whole Milk Powder was
mixed with calcite to create the spiked milk powder samples. Several samples with differing concentrations of calcium carbonate
were prepared by geometric dilution. Table 1 illustrates this process. Samples were obtained which contained 10%, 5.0%, 2.5%, 2.0%,
1.5%, 1.0% and 0.50% calcium carbonate concentration (weight by
weight; w/w). Three sets of these seven concentrations were prepared in order to monitor the consistency with which the samples
were being prepared and were labelled sample sets 1, 2 and 3 (SS1,
SS2 and SS3).
2.2. FT-Raman spectroscopy
FT-Raman spectra were recorded using an Equinox 55 interferometer bench (Bruker Optics, Ettlingen, Germany) equipped with
a FRA-106 Raman accessory and a D418-T liquid nitrogen cooled
Ge detector. A 1064 nm Nd:YAG laser was used to generate Raman
scattering. The FT-Raman spectrometer was operated using OPUS
5.5 (Bruker Optics, Ettlingen, Germany).
Samples were packed by hand into a small divet (approximately 10 mg) and analysed using the vertical objective in a 180
backscattering arrangement. Three spectra were taken for each
sample, rotating the divet between each spectrum. The FT-Raman
spectrometer parameters used to analyse the spiked milk powder samples include: 1064 nm laser wavelength to generate Raman
scattering, resolution was 4 cm1 for all samples, laser power was
120 mW and each sample was scanned 64 times. Each of the three
sample sets were analysed in this way on three separate occasions
(usually on different days) in order to include the day to day variability. Sets of spectra which were collected from the same sample
set but on a different occasion will be referred to as repetition 1, 2 or
3 depending on whether the spectra were taken on the rst, second
or third occasion, respectively. Therefore, in total, there were 9 different sets of spectroscopic data produced. Each of these sets of data
were analysed using both univariate and multivariate methods.
2.3. Data analysis
Multivariate analysis of spectra was performed using The
Unscrambler X (CAMO Technologies Inc., Woodbridge, USA). Initially the spectra were scaled using standard normal variate to
remove the effects of baseline offset between spectra which might
be caused by variation in surface layer morphology and laser focus.
Principal component analysis (PCA) was then used to identify outliers and to observe how well each of the spiked milk powder
samples separated. Partial least squares (PLS) analysis was performed using the NIPALS algorithm to calculate the PLS components
and generate prediction models. Cross validation was utilised as the
validation method. The appropriate number of factors to use was
determined from the explained variance curve and the plot of root
mean square error of cross validation (RMSECV) versus the number of factors used. These methods have also been utilised in the
literature to determine the appropriate number of factors to use
[23,24].
Univariate analysis was also performed as a comparison to the
PLS method for creating a prediction model. This was performed
using three distinct methods to create the regressions necessary to
calculate LOD and LOQ. The rst method involved integrating the
area between 1066 and 1105 cm1 underneath the spectrum but
above a line connecting the two edges of the integration area and
using the resulting value to create the regressions (method A). The
second method involved using the ratio of the integrals found at

Fig. 1. Raman spectra of: (a) milk powder, (b) spiked milk powder and (c) calcite.

1085 and 1443 cm1 (integrated between 1416 and 1500 cm1 ) to
create the regressions (method B). The 1085 and 1443 cm1 peaks
were chosen because they are produced by calcite and milk powder,
respectively. The third method for producing regressions involved
using the ratio of absolute peak heights for the 1085 and 1443 cm1
bands (method C).
The capability of univariate and multivariate techniques to
model the collected data was assessed through calculation of RMSE
for PLS analysis and the standard error for univariate analysis.
Using these values, the limits of detection and quantication were
calculated and the use of each of these techniques with Raman spectroscopy for quantication of calcite in milk powder was evaluated.
3. Results and discussion
Calcium carbonate produces a prominent peak at 1085 cm1
which is still visible when analysing samples of milk powder spiked
with calcite. Typical spectra of milk powder (Fig. 1a), spiked milk
powder (Fig. 1b), and calcite (Fig. 1c) show that the 1085 cm1
calcite peak is distinctive in the spiked samples.
3.1. Principal component analysis
The purpose of running principal component analysis (PCA) in
this study was to identify outliers within the data and to observe
the distribution of samples spiked with differing quantities of
added calcite in the scores plot. PCA was only applied to the
9871178 cm1 region of the spectrum to ensure that irrelevant
regions would not affect the scores plot.
Fig. 2 contains a scores plot constructed for sample set 1 (SS1)
repetition 1. It is important to note that three scores plots were
collected for each sample set. The difference between the plots is
simply the occasion in which the spectra (used to construct the
plots) were collected. Due to each sample set and their respective
repetitions giving similar results, only SS1 repetition 1 will be presented here. In the scores plots obtained from PCA (including the
plot presented in Fig. 2) it is apparent that the samples containing
0%, 5% and 10% calcite (w/w) each form distinct clusters. Samples
containing 0.52.5% calcite (w/w), however, form one large cluster
making it difcult to distinguish between concentration differences
of only 0.5%. For example, in the scores plot of SS1 repetition 1, the
score values of the 1.5% and 2% samples overlap with respect to
both PC-1 and PC-2, thus making it difcult to conclude that these
two samples form separate clusters. These plots also show that pure

G.P.S. Smith et al. / Vibrational Spectroscopy 67 (2013) 8791

89

Table 1
The process by which each sample mixture was prepared.
X

Mass of X added (g)

Mass of MP added (g)

Product name

Concentration of CC in product (%, w/w)

CC
Product 1
Product 2
Product 3
Product 4
Product 5
Product 6
Product 7
Product 8
Product 7

1.0
2.0
4.0
8.0
1.0
1.0
1.0
0.50
0.50
0.60

1.0
2.0
4.0
2.0
1.0
1.0
0.25
0.50
0.50
0.20

Product 1
Product 2
Product 3
Product 4
Product 5
Product 6
Product 7
Product 8
Product 9
Product 10

50
25
12.5
10
5
2.5
2
1
0.5
1.5

MP, milk powder; CC, calcium carbonate.

peak integration value gives the highest percentage and standard

error and the lowest R2 values after averaging. Use of the ratio of
absolute peak intensities to create regressions gives the lowest
percentage and standard errors and the highest R2 values on average. LOD and LOQ were calculated using the average standard error
of the regression and the average slope. LOD is considered to be
(3 standard error)/slope. LOQ is considered to be (10 standard
error)/slope. This gives values with units of % (w/w) calcite. LOD
and LOQ for each method were found to be (respectively): 3.8%
and 13% for method A; 3.3% and 11% for method B; 1.4% and 4.7%
for method C. These results indicate that method C is the most
effective of the three for quantifying calcite. The least effective
is method A. The main assumption of calculating LOD and LOQ
using these techniques is that the results obtained from each of
the sample set repetitions are treated as though they are true
repetitions (i.e. that each repetition was collected on the same day
and each repetition came from the same sample set). A possible
source of uncertainty in the univariate results is the presence of
milk powder peaks at and around 1085 cm1 . This would make
the integrated area underneath the 1085 cm1 calcite peak less
reliable, thus resulting in a larger LOD and LOQ.

Fig. 2. Scores plot of SS1 repetition 1 including the pure milk powder data (0%). The
key on the right indicates the calcite content as a weight by weight (w/w) percentage.

milk powder (0% calcite) forms a cluster far from any other sample,
including the 0.5% calcite sample. This indicates that spiked milk
powder can be reliably distinguished from pure milk powder even
when very little calcite is added. There was only one outlier (other
than pure milk powder) identied using PCA which was present in
SS1 repetition 2.

3.3. Partial least squares analysis

Spectra used in PCA were also used for PLS except for an outlier
identied using PCA. An XY relationship plot was created from PLS
data which initially included the pure milk powder data; however,
this plot indicated that pure milk powder acted as an inuential
outlier. This is seen in the Fig. 3 as a group of data points which
behave signicantly different to all other data points and have a
large inuence on the gradient of the tted line. Therefore, pure
milk powder data were also excluded from the PLS model in order
to avoid this effect.

3.2. Univariate analysis

Univariate analysis was performed using three distinct methods which shall be referred to as methods A, B and C. Slopes, R2
values, percentage errors and standard errors for each univariate
method and their respective sample set repetitions are provided
in Table 2. The results in this table show that the use of a single

Table 2
The slopes, percentage errors of slopes, R2 values and the standard errors obtained from univariate methods A, B and C.
Samplea

% Errorb

Slope

R2

Standard errorc

SS1R1
SS1R2
SS1R3
SS2R1
SS2R2
SS2R3
SS3R1
SS3R2
SS3R3

0.0038
0.0068
0.0082
0.0089
0.0033
0.0050
0.0030
0.0043
0.0030

0.017
0.02
0.017
0.017
0.016
0.015
0.014
0.014
0.015

0.051
0.051
0.048
0.043
0.047
0.043
0.043
0.039
0.040

13
13
5.1
6.0
13
7.1
12
7.6
13

5.4
9.6
6.7
4.7
7.1
7.9
8.8
9.4
9.0

2.5
4.0
3.0
2.2
3.9
3.2
3.2
4.5
3.6

0.76
0.77
0.95
0.94
0.77
0.91
0.78
0.90
0.76

0.95
0.86
0.92
0.96
0.91
0.89
0.87
0.86
0.87

0.99
0.97
0.98
0.99
0.97
0.98
0.98
0.96
0.97

0.007
0.012
0.006
0.0075
0.0059
0.0050
0.0051
0.0046
0.0055

0.016
0.026
0.016
0.011
0.016
0.017
0.018
0.019
0.020

0.018
0.029
0.020
0.013
0.026
0.019
0.019
0.025
0.021

Average

0.0051

0.016

0.045

10

7.6

3.3

0.84

0.9

0.98

0.0065

0.018

0.021

a
b
c

Refers to the sample set (SS) and its respective repetition (R).
Refers to the percentage error of the slope and was calculated from the standard error of the slope.
Refers to the standard error of the regression and was used to calculate the LOD and LOQ. These values were not used to calculate the percentage error of the slope.

90

G.P.S. Smith et al. / Vibrational Spectroscopy 67 (2013) 8791

Fig. 3. (a) XY relationship outliers plot constructed with pure milk powder (0%) included. (b) XY relationship outliers plot constructed with pure milk powder data excluded.
The key in each plot indicates the calcite concentration (w/w) of each sample.

PLS loadings convey the relationship between the raw spectroscopic data and its scores and may be interpreted in a similar
fashion as in PCA [25]. Therefore the structure within the spectroscopic data matrix correlated to the dominant variation with
changing calcite content will have the largest coefcients in the
loadings plot. Regression coefcients in this plot describe vectors
which are used to calculate concentrations which are presented
as predicted concentrations in the predicted versus measured plot.
Fig. 4 shows a loadings plot of factor 1 and indicates that the Raman
shift at 1085 cm1 has the largest number of regression coefcients.
This suggests that the peak at 1085 cm1 has the greatest amount
of inuence when calibrating the predicted versus measured concentrations plot and hence is the most useful for quantifying calcite
when using factor 1.
The predicted versus measured plot was utilised as the prediction model for multivariate analysis of spiked milk powder samples.
This plot illustrates the relationship between measured concentration of calcite with the concentrations predicted to be in the
samples by PLS. Full cross validation was utilised as the validation
method. The optimum number of factors to use to calibrate this plot
was determined from a plot of RMSE versus the number of factors
and an explained variance curve after cross validation (Fig. 5a and
b, respectively). The explained variance plot indicates that 99% of
variance is explained by factor 1 and factor 2 explains the remaining

variance. The RMSE versus number of factors plot shows little difference in RMSECV between factor 1 and the remaining factors. One
factor was, therefore, considered to be the optimum number to use.
Fig. 5b shows the predicted versus measured plot of SS1 repetition
1. The slope of this plot is 0.99 and the R2 value is 0.99, suggesting
that calcite content was able to be effectively predicted using PLS
and the plotted line ts well. The slopes, R2 values and the errors
of all sample sets are summarised in Table 3. The average slope
obtained from all samples is 0.99. The errors of the predicted concentrations are the RMSEC, RMSEP and RMSECV values. The average
RMSEC, RMSEP and RMSECV values obtained are 0.34, 0.38 and
0.52% (w/w), respectively. It should also be noted that the calculated error values are similar to error values reported in a study
of Raman spectroscopic quantication of fat and protein in milk
powder using PLS [6]. The average LOD and LOQ values for PLS are
1.0% and 3.4%, respectively. The values were calculated using the
RMSE rather than standard error as explained in Section 3.2. This
assumes that there is little difference between RMSE and standard
error values. These results suggest that it is also possible to predict calcite content in dry-mixed calcite spiked commercial milk
powders using Raman spectroscopy and PLS.
3.4. Comparison of PLS with univariate analysis
Univariate analysis results suggest that the use of absolute peak
intensities at 1085 and 1443 cm1 is the most effective univariate
method for quantifying calcite. However, PLS appears to be more
Table 3
The slopes, RMSEC, RMSEP, RMSECV and R2 values obtained from the predicted
versus measured plots of all sample sets and their respective repetitions (occasions
on which data was collected).

Fig. 4. A loadings plot constructed from factor 1 of SS1 repetition 1 after PLS analysis.

Sample set

Occasion

Slope

RMSECa

RMSEPa

RMSECVa

R2

SS1

Rep 1
Rep 2
Rep 3

0.99
0.96
0.99

0.31
0.65
0.17

0.35
0.71
0.19

0.39
0.76
0.47

0.99
0.96
0.99

SS2

Rep 1
Rep 2
Rep 3

0.99
0.99
0.99

0.20
0.35
0.29

0.23
0.41
0.32

0.27
0.49
0.48

0.99
0.99
0.99

SS3

Rep 1
Rep 2
Rep 3

0.99
0.98
0.99

0.33
0.39
0.34

0.38
0.44
0.37

0.54
0.54
0.71

0.99
0.98
0.99

Average

N/A

0.99

0.34

0.38

0.52

0.99

Unit: %, w/w.

G.P.S. Smith et al. / Vibrational Spectroscopy 67 (2013) 8791

91

Fig. 5. (a) Plot of root mean square error versus the number of factors used for calibration of the predicted versus measured plot. (b) Predicted versus measured plot calibrated
using one factor.

effective than univariate method C for this purpose. PLS gave LOD
and LOQ values of 1.0% and 3.4%, respectively, while method C gave
1.4% and 4.7%. PLS also produced more consistent gradients and R2
values from the standard curves it produced (the predicted versus
measured plots). However, discussions with a manufacturer of calcium fortied milk powder have conrmed that it typically contains
less than 3.2% (w/w) calcite. This suggests that it may not be possible to effectively quantify calcite content in retail milk powder
using PLS or univariate analysis. The errors calculated for this technique are similar to those found in the literature for the prediction
of fat and protein content in milk powder using the same technique,
which suggests that the calculated errors are accurate. Despite having a larger LOQ than the concentration of calcite typically present
in retail milk powder, the LOD from PLS analysis and univariate
method C are considerably lower than 3.2%. This indicates that it
is possible to detect calcite addition in some retail milk powder
samples using these techniques.
4. Conclusions
Principal component analysis can be used to identify spiked
milk powder samples as all samples with added calcite form clusters distinct from pure milk powder clusters. Samples with larger
differences in calcite content were also able to be distinguished
effectively. However, samples with concentration differences of
only 0.5% were difcult to separate by PCA. Investigation of PLS
and univariate analysis indicates that both of these techniques can
be used with Raman spectroscopic data to quantify calcite within
milk powder samples. The LOD and LOQ values obtained using each
technique show that PLS is the more effective technique for doing
so. Nevertheless, the PLS LOQ is larger than the quantity of calcite
typically used to fortify retail milk powder, indicating that PLS may
not be useful for quantifying calcite concentrations as low as 3.2%
(w/w). The calculated LOD values, however, indicate that PLS and
univariate method C could still be used to detect the presence of
calcite in such milk powder samples.
Acknowledgements
I would like to thank Austina Clark for her assistance with the
statistics involved in this investigation. I would also like to thank

Elizabeth Nickless for her miscellaneous feedback during this investigation.

References
[1] Index Mundi, Whole Milk Powder Domestic Consumption by Country
(accessed 30.07.12), sourced from the US Dept. of Agriculture, 2012
http://www.indexmundi.com/agriculture/?commodity=powdered-wholemilk&graph=domestic-consumption
[2] M. Caric, Concentrated and Dried Dairy Products, rst ed., VCH Publishers, Inc.,
New York, 1994.
[3] H.M. Barker, Nutrition and Dietetics for Health Care, tenth ed., Elsevier Health
Sciences, Philadelphia, 2002.
[4] D.D. Miller, Adv. Food Nutr. Res. 33 (1989) 103156.
[5] H.K. Vyas, P.S. Tong, J. Dairy Sci. 87 (2004) 11771180.
[6] C.M. McGoverin, A.S.S. Clark, S.E. Holroyd, K.C. Gordon, Anal. Chim. Acta 673
(2010) 2632.
[7] D. Wu, S. Feng, Y. He, J. Dairy Sci. 90 (2007) 36133619.
[8] D. Wu, Y. He, S. Feng, Anal. Chim. Acta 610 (2008) 232242.
[9] D. Wu, Y. He, J. Shi, S. Feng, J. Agric. Food Chem. 57 (2009) 16971704.
[10] D. Wu, Y. He, S. Feng, D. Sun, J. Food Eng. 84 (2008) 124131.
[11] A. Borin, M.F. Ferrao, C. Mello, D.A. Maretto, R.J. Poppi, Anal. Chim. Acta 579
(2006) 2532.
[12] H. Patel, S.G. Anema, S.E. Holroyd, H. Singh, L.K. Creamer, Lait 87 (2007)
251268.
[13] C. Reh, S.N. Bhat, S. Berrut, Food Chem. 86 (2004) 457464.
[14] S. Gallier, K.C. Gordon, R. Jimnez-Flores, D.W. Everett, Int. Dairy J. 21 (2011)
402412.
[15] C.M. McGoverin, A.S.S. Clark, S.E. Holroyd, K.C. Gordon, Anal. Method 1 (2009)
2938.
[16] R.M. El-Abassy, P.J. Eravuchira, P. Donfack, B. von der Kammer, A. Materny, Vib.
Spectrosc. 56 (2011) 38.
[17] M.R. Almeida, K.S. Oliveira, R. Stephani, L.F.C. de Oliveira, J. Raman Spectrosc.
42 (2011) 15481552.
[18] B. Kim, L.B. Perkins, R.J. Bushway, S. Nesbit, T. Fan, R. Sheridan, V. Greene, J.
AOAC Int. 91 (2008) 408413.
[19] E.C.Y. Li-Chan, Trends Food Sci. Technol. 7 (1996) 361370.
[20] U.P. Agarwal, R.S. Reiner, S.A. Ralph, Cellulose 17 (2010) 721733.
[21] E. Katainen, M. Elomaa, U. Laakkonen, E. Sippola, P. Niemela, J. Suhonen, K.
Jrvinen, J. Forensic Sci. 52 (2007) 8892.
[22] E.A. Sagitova, P. Donfack, K.A. Prokhorov, G.Y. Nikolaeva, V.A. Gerasin, N.D.
Merekalova, A. Materny, E.M. Antipov, P.P. Pashinin, J. Phys. Chem. B 116 (2012)
221231.
[23] C.M. McGoverin, Raman Spectroscopy of Complex Mixtures, University of
Otago, 2008 (Thesis (PhD)).
[24] K.R. Beebe, R.J. Pell, M.B. Seasholtz, Chemometrics A Practical Guide, rst ed.,
John Wiley and Sons, New York, 1998.
[25] K.H. Esbensen, Multivariate Data Analysis in Practice, fth ed., CAMO Process
AS, Oslo, 2002.