Food Chemistry 136 (2013) 1444–1452

Contents lists available at SciVerse ScienceDirect

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

Analytical Methods

Rapid analysis of sugars in honey by processing Raman spectrum using

chemometric methods and artificial neural networks
_
Beril Özbalci a, Ismail Hakkı Boyaci a,⇑, Ali Topcu a, Cem Kadılar b, Uğur Tamer c
a
Department of Food Engineering, Faculty of Engineering, Hacettepe University, 06800 Beytepe Ankara, Turkey
b
Department of Statistics, Faculty of Science, Hacettepe University, 06800 Beytepe Ankara, Turkey
c
Department of Analytical Chemistry, Faculty of Pharmacy, Gazi University, 06330 Ankara, Turkey

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

Article history: The aim of this study was to quantify glucose, fructose, sucrose and maltose contents of honey samples
Received 9 February 2012 using Raman spectroscopy as a rapid method. By performing a single measurement, quantifications of
Received in revised form 3 August 2012 sugar contents have been said to be unaffordable according to the molecular similarities between sugar
Accepted 16 September 2012
molecules in honey matrix. This bottleneck was overcome by coupling Raman spectroscopy with chemo-
Available online 28 September 2012
metric methods (principal component analysis (PCA) and partial least squares (PLS)) and an artificial neu-
ral network (ANN). Model solutions of four sugars were processed with PCA and significant separation
Keywords:
was observed. This operation, done with the spectral features by using PLS and ANN methods, led to
Honey
Glucose
the discriminant analysis of sugar contents. Models/trained networks were created using a calibration
Fructose data set and evaluated using a validation data set. The correlation coefficient values between actual
Maltose and predicted values of glucose, fructose, sucrose and maltose were determined as 0.964, 0.965, 0.968
Sucrose and 0.949 for PLS and 0.965, 0.965, 0.978 and 0.956 for ANN, respectively. The requirement of rapid anal-
Raman spectroscopy ysis of sugar contents of commercial honeys has been met by the data processed within this article.
Chemometrics Ó 2012 Elsevier Ltd. All rights reserved.
Artificial neural network

1. Introduction activity (Molina, 1989; White, 1978). Although compositional fac-

Honey is a natural, complex and liquid food product processed
by honeybees. This nutritious food is produced by blending the
sweetened sap collected from varying natural sources with meta-
bolic gastric juice and sputtering it into beehives made of wax to
be ripened. This initial sweet secretory mix is kept in a private
environment where climate fluctuation leads to moisture loss.
The nectar, gathered in hives, becomes dense, and the initial water
content in the range of 30–70%, depending on the plant source, de-
creases to a final value of 17–18%. Through this maturation period,
the enzymes of the gastric juice age the gathered mixture to a final
product, which has previously been reported in the literature to
contain water, monosaccharides (fructose and glucose), disaccha-
rides (sucrose and maltose), minerals, amino acids, proteins and
acids (Doner, 1977; Mesallam & El-Shaarawy, 1987).
The composition of honey varies depending on the region, sea-
son, variety of bee, the plant source of nectar, adulteration, etc. At
this point, classification and the quality control of such a product
gains great priority. The main quality factors for honey are given
by Codex Alimentarius as follows: moisture content, mineral con-
tent (ash), acidity, hydroxymethylfurfural content and diastase
⇑ Corresponding author. Tel.: +90 312 297 61 46; fax: +90 312 299 21 23.
_
E-mail address: ihb@hacettepe.edu.tr (I.H. Boyaci).
tors such as sugars, acids and ash were not considered to have pri-
mary importance for quality control, some investigators believe
them to be crucial (Singhal, Kulkarni, & Rege, 1997). Especially,
the profiles of predominant sugars such as glucose, fructose, su-
crose and maltose have been associated with a wide variety of sec-
ond-order quality properties such as viscosity, hygroscopy,
granulation and energy value (Ouchemouk, Schweitzer, Bey,
Djoudad-Kadji, & Louaileche, 2009). Glucose/water and fructose/
glucose ratios were listed as the main factors that characterise
the crystallisation of honey samples (Bonvehi, 1989). It has been
highlighted that the crystallisation of various honey samples based
on the glucose and fructose content varied by crystal size and the
rate of crystallisation, which leads to uncertainties for the honey
handlers at the processing stage (Bhandari, D’Arcy, & Kelly,
1999). From this aspect, crystallisation should also be considered
by the fact that it is an undesirable property in handling, process-
ing and marketing (Assil, 1991). The content of disaccharides
(mainly, maltose and sucrose) has been considered as a tool for
the characterisation of honey. The maltose content with a comple-
mentary of some other oligosaccharides was used to classify Span-
ish honeys and to differentiate Brazilian honeys from several
geographical regions (Da Costa Leite et al., 2000; Mateo &
BoschReig, 1997; Ouchemouk et al., 2009; Weston & Brocklebank,
1999). Also, the relative proportions of the main sugars contained

0308-8146/$ - see front matter Ó 2012 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.foodchem.2012.09.064
 B. Özbalci et al. / Food Chemistry 136 (2013) 1444–1452
in nectars (fructose, glucose and sucrose) were established to be
rather variable. However, they are quite consistent with certain
botanical families. The main role of sucrose was pointed out for
the characterisation of honey based on origin (Mateo & BoschReig,
1997).
For the determination of the sugar content of honey, many stud-
ies have been carried out using high-performance liquid chroma-
tography (HPLC) equipped with a differential refractive index
detector (Romero, Manzanares, García, Galdón, & Rodríguez,
2011), gas chromatography equipped with a flame ionisation
detector (Kaskoniene, Venskutonis, & Ceksteryte, 2011), nuclear
magnetic resonance spectroscopy (NMR), and Fourier transform
infrared spectroscopy (FTIR–ATR) (Justino et al., 1998). All the pro-
cedures listed above were considered to be time-consuming, not
cost effective and labour-intensive due to sample pre-treatment
and the need for expensive chemicals.
Nowadays, the Raman spectroscopy is increasingly used as an
analytical technique for the evaluation of food safety and quality.
The principle of Raman signals arises from the inelastic scattering
of the incident light from a sample and the frequency shift of the
scattered light shifts in a manner of characteristic molecular vibra-
tions (Kneipp, Kneipp, Itzkan, Dasari, & Feld, 1999). The general
advantages of Raman spectroscopy over other spectroscopic sys-
tems are the non-interference from water present in the sample
with the Raman measurement, ease of sampling and measurement,
and minimal fluorescence interference of sample matrix varying
from sample to sample. Combining Raman spectroscopy with che-
mometric methods and vibrational spectroscopy has enabled enor-
mous progression for both quantitative and qualitative
measurements of food components (Yang, Irudayaraj, & Paradkar,
2005).
Multivariate analysis is used in spectroscopy to extract informa-
tion from complex spectra containing overlapping regions (Iruday-
araj & Paradkar, 2002). The most commonly used multivariate
calibration methods are principle component analysis (PCA) and
partial least squares (PLS). These techniques are used to reveal
the differences between apparently similar spectra by using the
correlation between the collected data and the change made on
variables of interest. This way, a huge amount of collected data
on a spectrum can be reduced to principal components that repre-
sent the whole picture of the data (Irudayaraj & Sivakesava, 2001).
The success of these methods depends upon the choice of proper
spectral range, and the number of variables employed in the model
(Irudayaraj & Paradkar, 2002).
In the last decade, artificial neural networks (ANNs), known as
a knowledge-based approach, have been applied for the control
of food quality (Huang et al., 2010; Kilic, Boyaci, Koksel, &
Kusmenoglu, 2007). ANNs have been developed as generalisations
of mathematical models of the human cognition system or neural
biology (Fausett, 1994). Because of the outstanding utilities of
this approach, it is not necessary to use any threshold values for
inspection of the sample and there is no strict mathematical
equation, which causes difficulties in the analysis (Basß, Dudak, &
Boyacı, 2007).
The determination of glucose content in the beverage industry
(Delfino et al., 2011) and the discrimination of sugar additives in
honey as adulteration (Irudayaraj & Paradkar, 2002) by using Ra-
man instrumentation have been performed previously. Consider-
ing these studies on quantification by Raman spectroscopy, the
quality control of honey by determining the sugar contents seems
to be a promising method. From this point of view, the determina-
tions of glucose, fructose, sucrose and maltose contents of varying
honey samples are presented in this work. The spectra obtained
from 40 model mixtures and 90 honey samples were used. The col-
lected data, from both prepared sugar mixtures and honey solu-
tions, were processed through PCA, PLS and ANN to develop
1445
prediction models in an attempt to achieve multicomponent quan-
titative analysis in complex mixtures. The results were compared
with the results of HPLC as a reference method.
2. Experimental
2.1. Chemicals
High-purity glucose and fructose (>99%) were purchased from
Sigma–Aldrich (Taufkirchen, Germany) and used without any fur-
ther treatment. Biochemical-grade sucrose and maltose powder
was purchased from Acros Organics (Geel, Belgium). All solutions
were prepared with ultra-pure water (18 MX cm) to reach the de-
sired concentrations.
2.2. Samples
In the first part of the study, the 40 model solutions, each con-
sisting of different ratio of glucose, fructose, sucrose and maltose
mixtures, were prepared. The total sugar contents of each model
solution was 20% and separately in the range of 0–14%. Model solu-
tions were used for PCA evaluation to check possible sample
grouping of Raman spectra.
In the second part of the study, 22 different honey samples (14
flower and 10 pine honeys) of varying brands were purchased from
a Turkish market and a total number of 90 honey samples were
prepared by blending the different brands. To obtain higher bands
in Raman spectra the sugar concentrations were set to 10 times
higher than that of HPLC measurement. A total of 30 g of honey
was dissolved in a final volume of 100 mL for Raman, whereas
3 g was dissolved in a final volume of 100 mL for HPLC measure-
ments. The strong bands obtained with the 30% honey concentra-
tion used during Raman have thought to increase the prediction
capabilities of this method. Subsequently, these prepared solutions
were centrifuged at 18,000 rpm for 10 min at room temperature
and supernatants were used for measurements.
2.3. Instrumentation and sample analysis
DeltaNu Examiner Raman microscope (Deltanu Inc., Laramie,
WY, USA) with a 785 nm laser source, a motorised microscope
stage sample holder, and a CCD detector was used. Instrument
parameters were as follows: 200 lL solution in glass Raman cuv-
ette, 75 mW laser power, and 30 s acquisition. Spectra were ob-
tained in the range of 200–2000 cm1 at a resolution of 2 cm1.
Each sample was manipulated with the built-in ‘‘automatic base-
line correct’’ functions of the software.
Glucose, fructose, sucrose and maltose contents of the sample
were also determined by the HPLC method. SpectraSystem HPLC
(ThermoFinnigan Inc., CA, USA) integrated with an auto sampler
including temperature control for the column (SpectraSystem
AS3000), a degasser system (SpectraSystem SCM1000), a quater-
nary gradient pump (SpectraSystem P4000), a refractive index
detector (Shodex RI-101, Showa Denko, NY, USA), and a personal
computer with a software package for system control and data
acquisition (ChromQuest 4.2.34) were used for analyses. The anal-
ysis of glucose, fructose, and sucrose + maltose were performed
isocratically at 0.6 mL/min flow rate and at 80 °C with a
300  7.8 mm i.d. cation exchange column (Rezex RCM column,
Ca2+, 8 lm, Torrance, CA). Sucrose and maltose separation was
not successful and coeluted. So, Rezex RSO column (Ag+, 12 lm,
200  10 mm, Torrance, CA) was used for maltose determination,
and separation was achieved isocratically at a flow rate of
0.3 mL/min and at 80 °C. For both HPLC columns, deionised water
was used as mobile phase and injection volume was 20 lL. Quan-
tification of sugars was based on the external standard method.
1446 B. Özbalci et al. / Food Chemistry 136 (2013) 1444–1452
Retention time for different sugars, their detection columns, and
linear regression equations were given in Table 1S. Chromatograms
of the sugar standards and calibration curves were given in
Fig. 1S–5S. To determine the concentrations of the coeluting com-
pounds the following calculation was applied.
Sucrose : ½PASþM  ðC M  RFM Þ=RFs
where PAS+M = peak area (or height) of mix (sucrose + maltose) at Re-
zex RCM column, CM = concentration of maltose evaluated by using
Rezex RSO column, RFM = response factor for maltose at Rezex RCM
column, and RFS = response factor for sucrose at Rezex RCM column.
2.4. Multivariate data analysis
The simultaneous determination of specific analytes in complex
media is very difficult and requires handling huge amounts of data.
For the elimination of repetitious data, the multivariate data analy-
sis techniques (PCA, PLS and ANN) were applied for the determina-
tion of glucose, fructose, sucrose, and maltose contents in honey
using Raman spectrum of the each samples. Before processing the
data with chemometric methods, spectrum normalisation based
on the broad band at 1406 cm1 was applied to compensate the
gross differences in the spectral response that were caused by phys-
ical effects. Discriminations of sugar content using a chemometric
method (PCA) were investigated with the 40 model solutions. PCA
analysis was performed by using a statistical package (SPSS 16.0,
SPSS Inc., Chicago, IL, USA). Modelling the glucose, fructose, sucrose
and maltose contents of the honey (Y) as a function of the Raman
spectrum (X) of the sample was computed by the PLS method using
the PLS toolbox for MATLAB (Eigenvector Research Inc., Wenatchee,
WA, USA). The data-set-obtained honey samples were divided into
a calibration and a validation subset with 70 and 20 objects, respec-
tively. Each subset contained spectra from pure as well as mixed
samples. The characteristics of both sets are shown in Table 1.
The validity of the calibration was investigated by calculating the
values of root mean square error of calibration (RMSEC) and the
correlation coefficient (r). RMSEC refers to the calibration model
uncertainty. The smaller RMSEC value means a smaller uncertainty
in the developed model and a better calibration model. The valida-
tion was further investigated using the root mean square
error of prediction (RMSEP) for the prediction of samples. To
investigate the prediction capability of the ANN towards to the
PLS method, the prediction models for glucose, fructose, sucrose,
and maltose contents of the honey were also developed by using
the feed-forward back propagation ANN technique. Data subsets
of the same size used in the PLS method were processed in this
part of the study. The effects of the network training parameters
(number of neurons, transfer functions and training algorithms)
on the prediction capability of the trained networks were
investigated using a script coded and designed on MATLABÒ 7.0
(The MathWorks, Inc., Natick, MA). The script was explained in
Table 1
Sugar contents of the samples used as calibration and prediction data sets.
Sugar Number of Minimum Maximum
samples (%, w/v) (%, w/v)
Calibration set
Glucose 70 22.0 44.8
Fructose 70 27.0 47.6
Sucrose 70 0.9 3.9
Maltose 70 4.2 9.0
Prediction set
Glucose 20 25.1 36.2
Fructose 20 28.9 42.0
Sucrose 20 1.2 3.4
Maltose 20 5.2 8.8
our previous article (Kiliç, Bas, & Boyaci, 2009). The number of
neurons refers to number of neurons in the hidden layer (varying
from 0 to 10). Linear (purelin), log-sigmoid (logsig) and hyperbolic
tangent sigmoid (tansig) were investigated as transfer functions.
Finally, 10 training algorithms as gradient descent backpropagation,
gradient descent with momentum backpropagation, gradient des-
cent with adaptive learning rate backpropagation, resilient back-
propagation, Fletcher–Powell conjugate gradient backpropagation,
Powell–Beale conjugate gradient backpropagation, scaled conju-
gate gradient backpropagation, BFGS quasi-Newton backpropaga-
tion, one-step secant backpropagation and Levenberg–Marquardt
backpropagation were investigated. The script creates all possible
networks in the investigated network structure. A total of 93
different neural networks were created for one training algorithm.
Totally, 930 different neural networks were created and trained
for 10 training algorithms. ANN study was performed using the
Artificial Neural Networks Toolbox in MATLABÒ 7.0. After the
training stage had been completed, the processed networks were
tested and the performances were evaluated by determining mean
square errors (MSE) and correlation coefficients of both training
and testing processes. For the network which had minimum MSE
and maximum regression coefficient values between experimental
data, the estimated data was selected as the optimal for further
applications. Average absolute deviation (AAD) between predicted
and measured data were calculated to compare the capabilities of
the PLS and ANN methods for predicting the sugar content of the
honey samples. The equation used for calculation of the AAD is
given below.
(" # )
Xp
AAD ¼ ðjv i;mea  v i;pre j=ðv i;mea Þ =p  100
i¼1
where v i;mea is the measured (HPLC) value, v i;pre is the predicted va-
lue and p is the total number of the sample (p = 90).
3. Results and discussion
3.1. Raman spectra of sugars
The Raman spectra of the aqueous solutions of glucose, fructose,
sucrose, maltose and honey were collected as described in the
instrumentation section as shown in Fig. 1. There were strong
bands obtained for each sugar enabling the determination of spe-
cific fingerprint spectra. The spectra of glucose and fructose differ-
entiated clearly. However, it was quite difficult to identify sucrose
and maltose spectra. The disaccharide structures consisting of glu-
cose and fructose were generating bands approximately the same
wavenumber, with slight shifts due to the conformational changes
of the two monosaccharides as they form a bond in between. For
the quantification of each sugar in the complex mixture, processing
specific bands by height or area was not successful, especially for
the disaccharides. Sugars were mixed together in a model solution
as per the amounts found in honey. The bands were overlapping
each other, which made it impossible to calculate the band dimen-
sions for quantification.
Mean
(%, w/v)
3.1.1. Glucose spectrum
In the low-wavenumber range (200–500 cm1), the skeletal
33.4
37.3
vibrational motions of d(C–C–C), d(C–C–O), d(C–O) and s(C–C)
2.4 was the dominant (Barrett, 1981; Hineno, 1977; Korolevich,
6.6 Zhbankov, & Sivchik, 1990; Mathlouthi & Luu, 1980b; Wells &
Atalla, 1990). The bands at 317 and 341 cm1 had been assigned
30.7 to a d(C–C–C) vibration and an endocyclic d(C–C–O) ring mode,
35.5 respectively (Mathlouthi & Luu, 1980b). The band present at 415
2.3
and 437 cm1 was the d(C2–C1–O1) bending vibration in a-glucose
7.0
and b-glucose, respectively (Mathlouthi & Luu, 1980b). The bands
 B. Özbalci et al. / Food Chemistry 136 (2013) 1444–1452
Fig. 1. The Raman spectra of glucose, fructose, sucrose, maltose and honey.
at 776 and 790 cm1 in the spectrum of the aqueous solutions
were assigned to the m(C–C) and d(C1–H1) vibrations in a-glucose,
and the 838 and 856 cm1 bands was the same vibrations in
b-glucose (Mathlouthi & Luu, 1980b). In a study, Raman spectra of
fructose and glucose, the m(C–O), d(C–C–H), m(C–C) and d(C–C–O)
vibrations, suggested a contribution to the region in between 820
and 950 cm1 (Söderholm, Roos, Meinander, & Hotokka, 1999).
The spectral range between 1000 and 1200 cm1 could be
attributed to m(C–C) (Barrett, 1981; Hineno, 1977; Kacurakova &
Mathlouthi, 1996; Korolevich et al., 1990; Vasko, Blackwel, &
Koenig, 1972; Wells & Atalla, 1990) and m(C–O) (Barrett, 1981;
Hineno, 1977; Korolevich et al., 1990; Mathlouthi & Luu, 1980b;
Vasko et al., 1972; Wells & Atalla, 1990), stretching vibrations
(Hineno, 1977; Kacurakova & Mathlouthi, 1996; Mathlouthi &
Luu, 1980b; Vasko et al., 1972). Smaller contributions from bend-
ing motions, such as d(C–C–O), d(C–C–C) and d(C–C–H) have also
been suggested (Söderholm et al., 1999). The vibrational bands at
976 and 1028 cm1 in the spectrum of glucose were assigned to
the m(C–O) vibration in the two anomers (Mathlouthi & Luu,
1980b). The band centred at 1106 cm1 could be assigned to the
d(C–O–C) angle-bending model (Kacurakova & Mathlouthi, 1996).
3.1.2. Fructose spectrum
It has been proposed that fructose corresponds with the bands
at 314 and 353 cm1, originating from the d(C–C–C) ring vibration
in the pyranoid and furanoid forms of fructose (Mathlouthi & Luu,
1980a). The band present at 419 cm1 could be originated from the
d(C–C–O) ring vibration in the pyranoid ring. We found that the
d(C–C–O) vibration was the most dominant contributor in this
spectral range (Hineno, 1977; Huvenne, Vergoten, Fleury, &
Legrand, 1981; Mathlouthi & Luu, 1980a). The band at 744 cm1
in fructopyranose and at 800 cm1 in fructofuranose could be
attributed to the m(C–C) vibration (Mathlouthi & Luu, 1980a).
Furthermore, Wells and Atalla (1990) suggested contributions of
R–HCH stretching vibrations in this region. It is established that
m(C–O), m(C–C) and d(C–C–H) vibrations contribute to the spectral
area mentioned above (Söderholm et al., 1999). The two bands at
911 and 933 cm1 in the spectrum of fructose have been assigned
to d(COH) and m(C–O) out-of-ring vibrations by Mathlouthi and Luu
(1980a). They also suggested that the higher bands at 1028 and
1054 cm1 could be assigned to the m(C–O) vibration in the pyra-
noid and furanoid rings.
1447
3.1.3. Sucrose spectrum
It has been proposed that the bands present in the spectrum of
fructose at 314 and 357 cm1 are originated from the d(C–C–C) ring
vibration in the pyranoid and furanoid forms of fructose ring
(Mathlouthi & Luu, 1980b). The 347 cm1 band can be attributed
to an endocyclic d(C–C–O) ring mode of the glucose ring. The
d(C–C–O) ring vibration in the pyranoid ring of fructose originates
the band present at 419 cm1 and dominates the spectral range of
300–500 cm1 (Hineno, 1977; Huvenne et al., 1981; Mathlouthi &
Luu, 1980b). The band at 544 cm1 was assumed to be originated
from the a-glycosidic bond of C1 on the glucosyl subunit which
is also present in the maltose spectrum. At the fructose spectrum,
the two bands at 744 and 800 cm1 might have arisen mainly from
the m(C–C) vibration of fructopyranose and fructofuranose, respec-
tively, which have shifted as a result of the glycosidic bond with a
glucose ring (Mathlouthi & Luu, 1980b). The band at 1028 cm1 in
both the maltose and the sucrose spectra has been assigned to the
m(C–O) vibration of the glucose ring.
3.1.4. Maltose spectrum
The 341 cm1 band could be assigned to an endocyclic d(C–C–O)
ring mode. The d(C2–C1–O1) bending vibration of D-glucose was at
434 cm1 (Mathlouthi & Luu, 1980b). The band at 544 cm1 was as-
sumed to be originated from the a-glycosidic bond of C1 on the glu-
cosyl subunit, which also presented in the sucrose spectrum. The
bands at 776 and 790 cm1 in the glucose spectrum of the aqueous
solutions that were assigned to the m(C–C) and d(C1–H1) vibrations
of a-glucose were widened and weakened in the maltose spectrum
as the bands seen at 842 and 855 cm1. The band at 1028 cm1 in
the spectrum of maltose and sucrose have been assigned to the
m(C–O) vibration of the glucose ring.
3.1.5. Honey spectrum
On the spectrum of honey the main bands at 314, 341, 415, 530,
617, 744, 776, 790, 838 cm1, and 856, 911, 933, 1028 and 1106 cm-1
were found to be in correlation with sugars. The observation of
glucose and fructose bands, the most on the honey spectrum, was
associated with high contents of these two compounds, whereas
sucrose and maltose bands have been suppressed. The band at
314 cm1, arising from fructose that originated from the d(C–C–C)
ring vibration in the pyranoid, has overlapped with the band of
glucose at 317 cm1, thus causing a broadening. Whereas the band
1448 B. Özbalci et al. / Food Chemistry 136 (2013) 1444–1452

on the fructose spectrum have overlapped. The strongest band at

530 and 616 cm1 has been clearly assigned to the fructose
spectrum, as it has the highest content. The bands at 744 and
at 800 cm1 were associated with the m(C–C) vibration in the
fructopyranose and fructofuranose forms of fructose, respectively
(Mathlouthi & Luu, 1980a). The peaks at 776, 790, 838 and
856 cm1 have been observed as being quite weak compared to their
appearances on the glucose spectrum. First they have been over-
lapped with the fructose bands, whereas the other two have been
broadened depending on the sample matrix. The two bands at 911
and 933 cm1 in the spectrum of fructose have been assigned to
d(COH) and m(C–O) out-of-ring vibrations by Mathlouthi and Luu
(1980a), and have been shortened according to concentration
changes and food matrix. The band at 1028 cm1 is thought to be
originated from the m(C–O) vibration of the glucose ring seen in
glucose, maltose and sucrose spectra. The shortening of the band
centred at 1106 cm1 could be assigned to the d(C–O–C) angle-
bending model (Kacurakova & Mathlouthi, 1996) of glucose.
Fig. 2. PCA plot for classification of sugars in standard solutions obtained from
Raman spectra in the region 200–2000 cm1.
3.8. Data mining

at 341 cm1 representing an endocyclic d(C–C–O) ring mode of 3.8.1. PCA

glucose has covered the band at 353 cm1, which originated from Principle component analysis is a suitable statistical method
the d(C–C–C) ring vibration in the furanoid form of fructose. The when there are many observed variables. A smaller number of arti-
bands 415 and 437 cm1 on the glucose spectrum and 419 cm1 ficial variables (called principle components), which will account

Fig. 3. The correlation between the number of principal components with RMSEC and RMSEC values in PLS calibration model for glucose, fructose, sucrose and maltose in
honey.
 B. Özbalci et al. / Food Chemistry 136 (2013) 1444–1452 1449

Fig. 4. Comparison of the sugar results obtained from conventional HPLC method with that of the PLS method for real samples. (A) Calibration data set and (B) prediction data
set.

for most of the variance in the observed variables, can be devel- predictor in subsequent analyses. By reviewing the correlations
oped by PCA. Then, these principal components may be used as a between the variables and the components, and using this
1450 B. Özbalci et al. / Food Chemistry 136 (2013) 1444–1452
Table 2
Selected network structures with the highest r value.
Sugar Training algorithms
Glucose Levenberg–Marquardt backpropagation
Fructose One step secant backpropagation
Sucrose Scaled conjugate gradient backpropagation
Maltose Levenberg–Marquardt backpropagation
r, correlation coefficient.
information, it is possible to interpret the components; that is, to
determine what construct seems to be measured by component
1, what construct seems to be measured by component 2, and so
forth. However, analysis of more than one component is difficult
for interpreting the unrotated factor pattern used in the statistical
methods. To make interpretation easier, it is necessary to perform
an operation called a rotation. A rotation is a linear transformation
that is performed on the factor solution for the purpose of making
the solution easier to interpret. Varimax rotation is probably the
most commonly used rotation. A varimax rotation is an orthogonal
rotation, meaning that it results in uncorrelated components. Com-
pared to some other types of rotations, a varimax rotation tends to
maximise the variance of a column of the factor pattern matrix.
Interpreting a rotated solution means determining just what is
measured by each of the retained components. Briefly, this in-
volves identifying the variables that demonstrate high loadings
for a given component, and determining what these variables have
in common. Usually, a brief name is assigned to each retained com-
ponent that describes its content.
In the first part of the study, all 40 model samples with different
ratios of glucose, fructose, sucrose and maltose were evaluated
by PCA to check possible sample grouping considering the
complete Raman spectra. By a varimax rotation, PCA resulted in
four principal components explaining 98.7% of the total variance.
Two-dimensional (2D) PCA graphs were plotted for the binary
combination of the four principal components. All graphs were
given in Fig. 6S. As it can be seen in the plot of the fourth versus first
principal components, each sugar was separated and can be clearly
distinguished in Fig. 2. This produces the conclusion of samples
arranged on the graph depending on their sugar contents and
specific locations dominated by the concentration of each sugar.
These lead up to the possibility of quantifying sugar concentrations
within the samples using chemometric methods.
3.8.2. PLS
The full region (200–2000 cm1) of the sample Raman spectrum
was used for calibration and validation of the PLS models. The ef-
fects of principal components on the RMSEC and RMSEP values
were visualised in Fig. 3. Taking into account RMSEC and RMSEP
values, the best principal component numbers were selected as
9, 7, 8 and 10 for glucose, fructose, sucrose and maltose predic-
tions, separately. Fig. 4 shows the PLS prediction results of glucose,
fructose, sucrose and maltose in honey for calibration data set (A)
and validation data set (B). The relationships between measured
(x-axis) and predicted (y-axis) results were defined with a linear
model (regression equation), of which the intercept value was zero.
The slopes of regression equations for all sugars investigated in this
work were in the range of 0.996 and 1.010. These results represent
the high prediction ability of PLS methods. These results were also
supported with high r values of the regression equations, found be-
tween 0.949 and 0.989.
3.8.3. ANN
The script coded for the optimisation was applied to glucose,
fructose, sucrose and maltose data sets. For each data set, 930 dif-
Hidden Neurons Transfer Function r Training data r Test data
#1 #1
5 Purelin Purelin 0.978 0.965
6 Purelin Purelin 0.972 0.965
9 Tansig Purelin 0.993 0.978
7 Purelin Purelin 0.978 0.956
ferent network topologies were created, trained and tested in order
to find the best neural network structure. During the training stage,
a validation data set was used for early stopping. Predicted outputs
of the train, validation, and test data sets were compared with the
actual outputs after the training stage of each single network and
subsequently the r and MSE values were calculated. After the train-
ing of all created topologies was completed, the results were col-
lected and combined within a file for investigation of the optimal
network structure(s). The networks that have the highest scores
for predictions of glucose, fructose, sucrose and maltose were given
in Table 2. In Fig. 5, plots of experimental data versus the data esti-
mated were given for the selected networks. Four networks were
the best performing among the 930 networks. However, if the fig-
ures are investigated in detail, the slopes of regression equations
were close to ‘‘1’’. Moreover, the intercepts of the graphs were
‘‘0’’. Correlation coefficient values of regression equations for glu-
cose, fructose, sucrose and maltose were in the range of 0.972
and 0.993 for calibration data sets and 0.956–0.978 for validation
data sets. Based on these results, it is possible to conclude that
the trained networks can be used successfully for the estimation
of sugar contents of honey using Raman spectra of the samples.
3.9. Comparison of the multivariate data analysis methods
The glucose, fructose, sucrose and maltose content of 90 honey
samples were determined in the range of 25.1–36.2%, 28.9–42.0%,
1.2–3.4% and 5.2–8.8% in compliance with the literature. To evalu-
ate the efficiencies of the two multivariate techniques (PLS and
ANN) in conjunction with Raman spectroscopy, commercial honey
samples were used for both calibration and validation sets. As a re-
sult, both techniques were found to be highly successful. The corre-
lation coefficients for glucose, fructose, sucrose and maltose were
very high and were calculated as 0.964, 0.965, 0.968 and 0.949 for
PLS and 0.965, 0.965, 0.978 and 0.956 for ANN, using Raman spec-
tra. For the comparison of prediction capabilities of the methods,
AAD values between measured and predicted values were also
investigated. AAD values of glucose, fructose, sucrose and maltose
in commercial honey samples were predicted using Raman spectra
as 1.9%, 3.0%, 3.5% and 2.8% for PLS and 1.9%, 2.0%, 3.2% and 2.7% for
ANN, respectively. These error values are thought to be originated
from the complex food matrix that cannot be overcome with the
use of other techniques. To compare with chromatographic tech-
niques, the determination of sugar content for the four sugars of
honey was achieved with a 2-min data collection, and with no
use of laborious equipment and, expensive chemicals.
As compared to the previous work (Batsoulis et al., 2005), the
prediction efficiency has been tripled by using a sufficient number
of training samples, and the reliability of the training was demon-
strated by predicting the sugar contents of 20 honey samples. The
estimation of four sugar content of honey was also achieved by
FTIR and chemometric methods previously with high regression
coefficients (Wang, Kliks, Jun, Jackson, & Li, 2010). However, the
efficiency of the calibration was evaluated by using model, syn-
thetic solutions. We have carried the analysis a step forward by
using commercial honeys for the calibration data set and involving
 B. Özbalci et al. / Food Chemistry 136 (2013) 1444–1452 1451

Fig. 5. Comparison of the sugar results obtained from conventional HPLC method with that of the ANN method for real samples. (A) Calibration data set and (B) prediction
data set.

the ANN method. The additional use of the ANN method provides work to a next level due to Raman’s attractive capability for obtain-
extra straightness for the prediction. Two multivariate techniques ing fingerprint signatures of all kinds of molecules, rapid-response
(PLS and ANN) with the Raman instrumentation have carried the times and the ability to work with biological samples. Raman has
1452 B. Özbalci et al. / Food Chemistry 136 (2013) 1444–1452
gone far beyond being a complementary tool like the molecular
vibration technique, as it provides direct molecular information
and conformational variations. The results obtained are clearly
preferable by error values and accuracy subjected to the amount
of data collected.
4. Conclusion
Glucose, fructose, sucrose and maltose contents in aqueous
solutions and in commercial honeys were determined by Raman
spectroscopy and by applying PCA, PLS and ANN methods for anal-
ysis of the spectral data. With the ease of sample preparation and
high-throughput multivariate methods, the quantification of four
sugars with a single measurement was achieved. The statistical
methods benefitted by avoiding the single-band process for every
single member of the training group, resulting in efficient quanti-
fication of the components individually without influencing each
other. The prediction capabilities of the techniques, PLS and ANN,
were tested, and it was observed that the selected models obtained
by PLS and ANN analyses are in close proximity from a statistical
point of view. These results demonstrate that Raman spectroscopy
in combination with appropriate multivariate methods can be suc-
cessfully adopted to quantitatively determine glucose, fructose, su-
crose and maltose contents in honey samples without any
treatment or chromatographic methods.
Acknowledgements
The authors are grateful for the financial support provided by
The Scientific and Technological Research Council of Turkey; Pro-
ject No.: 108O031.
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.2012.
09.064.
References
Assil, H. I. (1991). Crystal control in processed liquid honey. Journal of Food Science,
56(4), 1034.
Barrett, T. W. (1981). Laser Raman spectra of mono-, oligo- and polysaccharides in
solution. Spectrochimica Acta Part A: Molecular Spectroscopy, 37(4), 233–239.
Basß, D., Dudak, F. C., & Boyacı, I. H. (2007). Modeling and optimization III: Reaction
rate estimation using artificial neural network (ANN) without a kinetic model.
Journal of Food Engineering, 79(2), 622–628.
Batsoulis, A. N., Siatis, N. G., Kimbaris, A. C., Alissandrakis, E. K., Pappas, C. S.,
Tarantilis, P. A., et al. (2005). FT-Raman spectroscopic simultaneous
determination of fructose and glucose in honey. Journal of Agricultural and
Food Chemistry, 53(2), 207–210.
Bhandari, B., D’Arcy, B., & Kelly, C. (1999). Rheology and crystallization kinetics of
honey: Present status. International Journal of Food Properties, 2(3), 217–226.
Bonvehi, J. S. ( (1989). Physico-chemical properties. Composition and pollen
spectrum of eucalyptus honey produced in Spain. Anales de Bromatologia,
41(1), 41–56.
Da Costa Leite, J. M., Trugo, L. C., Costa, L. S. M., Quinteiro, L. M. C., Barth, O. M.,
Dutra, V. M. L., et al. (2000). Determination of oligosaccharides in Brazilian
honeys of different botanical origin. Food Chemistry, 70(1), 93–98.
Delfino, I., Camerlingo, C., Portaccio, M., Ventura, B. D., Mita, L., Mita, D. G., et al.
(2011). Visible micro-Raman spectroscopy for determining glucose content in
beverage industry. Food Chemistry, 127(2), 735–742.
Doner, L. W. (1977). The sugars of honey—A review. Journal of the Science of Food and
Agriculture, 28(5), 443–456.
Fausett, L. V. (1994). Fundamentals neural networks: Architectures, algorithms and
applications. New Jersey: Prentice Hall.
Hineno, M. (1977). IR-spectra and normal vibrations of beta-D-glucopyranose.
Carbohydrate Research, 56(2), 219–227.
Huang, Y., Lan, Y., Thomson, S. J., Fang, A., Hoffmann, W. C., & Laceyd, R. E. (2010).
Development of soft computing and applications in agricultural and biological
engineering. Computers and Electronics in Agriculture, 71(2), 107–127.
Huvenne, J. P., Vergoten, G., Fleury, G., & Legrand, P. (1981). Champ de forces de
symetrie locale de molecules cycliques: III. Interprétation des spectres de
vibration du glucose cristallisé sous les deux formes a et b. Journal of Molecular
Structure, 74(2), 169–180.
Irudayaraj, J., & Paradkar, M. M. (2002). Discrimination and classification of beet and
cane inverts in honey by FT-Raman spectroscopy. Food Chemistry, 76(2),
231–239.
Irudayaraj, J., & Sivakesava, S. (2001). Detection of inverted beet sugar adulteration
of honey by FTIR spectroscopy. Journal of Science of Food and Agriculture, 81,
683–690.
Justino, L. G., Caldeira, M., Gil, V. M. S., Baptista, M. T., da Cunha, A. P., & Gil, A. M.
(1998). Determination of changes in sugar composition during the aging of
honey by HPLC, FTIR and NMR spectroscopy. Carbohydrate Polymers, 34(4),
431–435.
Kacurakova, M., & Mathlouthi, M. (1996). FTIR and laser-Raman spectra of
oligosaccharides in water: Characterization of the glycosidic bond.
Carbohydrate Research, 284(2), 145–157.
Kaskoniene, V., Venskutonis, P. R., & Ceksteryte, V. (2011). Sugar analysis for
authenticity evaluation of honey in lithuanian market. Acta Alimentaria, 40(2),
205–216.
Kilic, K., Boyaci, I. H., Koksel, H., & Kusmenoglu, I. (2007). A classification system for
beans using computer vision system and artificial neural networks. Journal of
Food Engineering, 78(3), 897–904.
Kiliç, K., Bas, D., & Boyaci, I. H. (2009). An easy approach for the selection of optimal
neural network structure. GIDA, 34(2), 73–81.
Kneipp, K., Kneipp, H., Itzkan, I., Dasari, R. R., & Feld, M. S. (1999). Ultrasensitive
chemical analysis by Raman spectroscopy. Chemical Reviews, 99(10),
2957–2976.
Korolevich, M. V., Zhbankov, R. G., & Sivchik, V. V. (1990). Calculation of absorption-
band frequencies and intensities in the IR-spectrum of alpha-D-glucose in a
cluster. Journal of Molecular Structure, 220, 301–313.
Mateo, R., & BoschReig, F. (1997). Sugar profiles of Spanish unifloral honeys. Food
Chemistry, 60(1), 33–41.
Mathlouthi, M., & Luu, D. V. (1980a). Laser-Raman spectra of D-fructose in aqueous-
solution. Carbohydrate Research, 78(2), 225–233.
Mathlouthi, M., & Luu, D. V. (1980b). Laser-Raman spectra of D-glucose and sucrose
in aqueous-solution. Carbohydrate Research, 81(2), 203–212.
Mesallam, A. S., & El-Shaarawy, M. I. (1987). Quality attributes of honey in Saudi
Arabia. Food Chemistry, 25(1), 1–11.
Molina, C. L. H. (1989). Honey quality analysis. Alimentos, 14(4), 55–60.
Ouchemouk, S., Schweitzer, P., Bey, M. S., Djoudad-Kadji, H., & Louaileche, H. (2009).
HPLC sugar profiles of Algerian honeys. Food Chemistry, 121, 561–568.
Romero, C. D., Manzanares, A. B., García, Z. H., Galdón, B. R., & Rodríguez, E. R.
(2011). Differentiation of blossom and honeydew honeys using multivariate
analysis on the physicochemical parameters and sugar composition. Food
Chemistry, 126, 664–672.
Singhal, R. S., Kulkarni, P. R., & Rege, D. V. (1997). Handbook of indices of food quality
and authenticity. Cambridge: Woodhead Publishing.
Söderholm, S., Roos, Y. H., Meinander, N., & Hotokka, M. (1999). Raman spectra of
fructose and glucose in the amorphous and crystalline states. Journal of Raman
Spectroscopy, 30(11), 1009–1018.
Vasko, P. D., Blackwel, J., & Koenig, J. L. (1972). Infrared and Raman spectroscopy of
carbohydrates. 2. Normal coordinate analysis of alpha-D-glucose. Carbohydrate
Research, 23(3), 407.
Wang, J., Kliks, M. M., Jun, S., Jackson, M., & Li, Q. X. (2010). Rapid analysis of glucose,
fructose, sucrose, and maltose in honeys from different geographic regions
using Fourier transform infrared spectroscopy and multivariate analysis. Journal
of Food Science, 75(2), 208–214.
Wells, H. A., & Atalla, R. H. (1990). An investigation of the vibrational-spectra of
glucose, galactose and mannose. Journal of Molecular Structure, 224, 385–424.
Weston, R. J., & Brocklebank, L. K. (1999). The oligosaccharide composition of some
New Zealand honeys. Food Chemistry, 64(1), 33–37.
White, J. W. (1978). Honey. Advances in Food Research, 24, 287–374.
Yang, H., Irudayaraj, J., & Paradkar, M. M. (2005). Discriminant analysis of edible oils
and fats by FTIR, FT-NIR and FT-Raman spectroscopy. Food Chemistry, 93, 25–32.