NIR Red Wines

7/27/2019 NIR Red Wines

1/13

Analytica Chimica Acta 701 (2011) 139151

Contents lists available at ScienceDirect

Analytica Chimica Acta

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

Adulteration ofthe anthocyanin content ofred wines: Perspectives forauthentication by Fourier Transform-Near InfraRed and 1H NMRspectroscopies

E. Ferrari a, G. Focab,, M. Vignali c, L. Tassi a, A. Ulrici b

a Dipartimento di Chimica, Universit diModena e Reggio Emilia, Via Campi 183, 41125Modena, Italyb Dipartimento di Scienze Agrarie e degli Alimenti, Universit di Modena e Reggio Emilia, Padiglione Besta, ViaAmendola 2, 42122 Reggio Emilia, Italyc Vinicola SanNazaro, ViaGonzaga 12, 46020 Pegognaga (MN), Italy

a r t i c l e i n f o

Article history:Received 24 March 2011Received in revised form 23 May 2011Accepted 26 May 2011Available online 22 June 2011

Keywords:

Wine adulterationAnthocyaninsFourier Transform-Near InfraRed1H NMRMultivariate classificationFeature selection

a b s t r a c t

In the Italian oenological industry, the regular practice used to naturally increase the colour ofred winesconsists in blending them with a wine very rich in anthocyanins, namely Rossissimo. In the Asian market,on the other hand, anthocyanins extracted by black rice are frequently used as correctors for wine colour.This practice does not produce negative effects on health; however, in many countries, it is consideredas a food adulteration.

The present study is therefore aimed to discriminate wines containing anthocyanins originated fromblack rice and grapevine by using reliable spectroscopic techniques requiring minimum sample prepara-tion. Two series ofsamples have been prepared from five original wines, that were added with differentamounts ofRossissimo or ofblack rice anthocyanins solution, until the desired Colour Index was reached.The samples have been analysed by FT-NIRand 1HNMRspectroscopies and the resulting spectra matriceswere subjected to multivariate classification. Initially, PLS-DA was used as classification method, thenalso variable selection/classification methods were applied, i.e. iPLS-DA and WILMA-D. The classificationwith variable selection of NIR spectra permitted to classify the test set samples with an efficiency ofabout 70%. Probably these not excellent performances are due to the matrix effect, together with the lackof sensitivity ofNIRwith respect to minor compounds. On the contrary, very satisfactory results were

obtained on NMR spectra in the aromatic region between 6.5 and 9.5 ppm. The classification methodbased on wavelet-based variables selection, permitted to reach an efficiency in validation greater than95%.

Finally, 2D correlation analysis was applied to FT-NIRand 1H NMRmatrices, in order to recognise thespectral zones bringing the same chemical information.

2011 Elsevier B.V. All rights reserved.

1. Introduction

The wine-making process for red wines production involvesthe extraction of different phenolic compounds. Among them, theanthocyanins are particularly relevant, because they are the mainlyresponsible substances forred wines colour. InVitisvinifera grapesand in the corresponding wines are present malvidin, peonidin,

delphinidin, petunidin and cyanidin the aglycone moieties ofanthocyanin molecules which differ from one another for thenumber of hydroxyl and methoxyl groups that can serve as sub-stituents of the aromatic rings. Such aglycones occur in nature inglycosylated forms with one or more sugars [1,2]. Anthocyaninsare present in Italian wines in widely varying amounts usu-ally 1001500mg L1 depending on the grape variety, various

Corresponding author. Tel.: +39 0522 522042; fax: +39 0522 522027.E-mail address: [email protected] (G. Foca).

seasonal and environmental factors as well as the techniques ofvinification and the length of the aging process [3].

In the Italian oenological industry, the common and standardpractice used to naturally increase the colour of red wines consistsin blending them with a wine very rich in anthocyanins, namelyRossissimodellEmiliaor, simply, Rossissimo [4]. In the Asian market,on the other hand, anthocyanins extracted by black rice are fre-

quently used as correctors for wine colour. This practice is boostedby the fact that the extraction of anthocyanins from black rice isvery advantageous, since the total content of anthocyanins in thewhole seed can reach up to 500 mg/100g [5], and anthocyanins aremainly located in the husk, that is usually removed as refuse. Suchas for the types of anthocyanins extracted by black rice, differentstudies have shown that the aglycone moieties mainly representedarecyanidinandpeonidin [58]. Sincetheanthocyaninsfamilycon-sists of molecules structurally very similar and of proven salutaryeffect, it is easy to understand that the problem of using antho-cyanins extracted from black rice to increase the colour of wineis not a problem of toxicity or of food safety, but it is a merely

0003-2670/$ see front matter 2011 Elsevier B.V. All rights reserved.

doi:10.1016/j.aca.2011.05.053
http://localhost/var/www/apps/conversion/tmp/scratch_2/dx.doi.org/10.1016/j.aca.2011.05.053http://localhost/var/www/apps/conversion/tmp/scratch_2/dx.doi.org/10.1016/j.aca.2011.05.053http://www.sciencedirect.com/science/journal/00032670http://www.elsevier.com/locate/acamailto:[email protected]://localhost/var/www/apps/conversion/tmp/scratch_2/dx.doi.org/10.1016/j.aca.2011.05.053http://localhost/var/www/apps/conversion/tmp/scratch_2/dx.doi.org/10.1016/j.aca.2011.05.053mailto:[email protected]://www.elsevier.com/locate/acahttp://www.sciencedirect.com/science/journal/00032670http://localhost/var/www/apps/conversion/tmp/scratch_2/dx.doi.org/10.1016/j.aca.2011.05.053


2/13

140 E. Ferrari et al. / Analytica Chimica Acta 701 (2011) 139151

legislative one. In any case, in many countries, Italy included, thisprocedure is considered in effect as a food adulteration.

The present study is therefore aimed at discriminating winescontaining anthocyanins added with Rossissimo from wines addedwith black rice solution.

Currently, the total anthocyanins content in wine usuallyexpressed as malvidin-3-glucoside equivalents is determined byUVvis analysis at 520 nm [9], that is definitely a rapid and reli-

able method, but it does not permit the speciation of the singleanthocyanins that is indeed necessary to identify their origin.The provenance of anthocyanins from plant species different

from grapevine, on the contrary, can be assessed by chromato-graphic techniques, as widely confirmed by literature. A lot ofworks, in fact, deal with the identification and quantification ofspecific anthocyanins by means of reverse-phase HPLC [1,2,913]or liquidliquid counter-current chromatography [14,15]. Liquidchromatography is a very sensitive technique but, unfortunately, itis rather time-consuming, because it usually needs a long time forthe preparation of samples and standard solutions, besides the factthat it requires skilled personnel.

In thewinemaking sector, at thepresent time, theinfraredspec-troscopies are widely used for routine analysis of musts and winessince they offer the advantage that they do not need sample prepa-ration. The infrared spectrum is, in fact, acquired on the complexmatrix in a way to obtain a sample fingerprint, then the simulta-neous determination of different analytes from the spectra is madeby applying multivariate calibration models, previously developedon a set of reference samples.

A number of applications for determiningthe total anthocyaninscontent in wine by NIR/visNIR can be found in literature, i.e. theworks by Bauer et al. [16], Janik et al. [17], Di Egidio et al. [18] andCozzolino et al. [19], who also demonstrated the feasibility of usingNIR spectroscopy to determine other grape phenolic compounds[20], while Sinelli et al. [21] and Zsivanovits et al. [22] predictedthe total anthocyanins content by NIR in berries.

Anothersuitabletechniqueforthecharacterisationoftheantho-cyaniccomponent of wines is NMRspectroscopy, sinceit shares the

main advantage of NIR spectroscopy, i.e. the possibility to analysethe sample as it is, but avoids its main defect, i.e. the scarce sen-sitivity in determining compounds present at low concentrations[23,24], such as anthocyanins, that are present in wine in amountsclose to theNIR limit of detection. With respect to chromatographicanalyses, NMRis a little less sensitive, but it is non-destructive,simpler and shorter as sample preparation, more reproducible andcapable of simultaneous detection of a great number of low molec-ular mass components in complex mixtures [2527].

In wine analysis context do exist some works about the appli-cation of NMRspectroscopy for the analysis of wine samples ascomplex matrices, i.e. without the need for sample preparation orisolation of analytes. Some authors have analysed wine sampleswithout extraction or purification, then they calculated from the

NMRsignals a certain numberof parameters to be used forbuildingthedatamatrix for chemometricanalysis[28,29]. Inadifferentway,most of the authors preferred to apply chemometric techniques onthe whole NMRspectra, presenting overlapping peaks as well, fol-lowing the so-called blind analysis of signals to face classificationor calibration problems [3033].

With the aim of discriminating wines containing anthocyaninsadded with Rossissimo from wines added with black rice solution,we have initially chosen to assess the applicability of the simplestand fastest technique, i.e. the NIR spectroscopy. The replicate NIRspectra of samples taken from the two classes were preprocessedand then analysed by means of PLS-DA in order to build the cor-responding classification model. Since the anthocyanins-relatedinformation could be overwhelmed by uninformative variation, it

was also applied a novel classification method of variable selection

in the wavelet domain, called WILMA-D (Wavelet Iterative LinearModelling Approach-Discriminant version). The aim ofWILMA-D introduced forthe first time in thepresent work is thediscrimina-tion between objects of differentclasses andit is a modified versionof the algorithm WILMA [34,35], which was originally devoted tosolving regression problems.

Theobtained results have shown that NIRspectroscopy seems toreveal some useful information to distinguish between adulteratedand non-adulterated samples, but the matrix effect together withthe lack of sensitivity of NIR with respect to minor compounds,hindered to obtain high performances in validation.

As a consequence, we also turned our attention to the 1H NMRspectroscopy, since it is a more sensitive technique but rather sim-ple as for sample preparation. After the alignment of spectra, thearomatic compound region between 6.5 and 9.5ppm was retainedfor the successive classification. Initially, PLS-DA was used as clas-sification method on signals preprocessed in different ways then,also in this case, the variable selection was attempted. In additionto WILMA-D, another methodof variable selection, i.e. iPLS-DA, wasinvestigated. Since the NMRspectra present less overlapped bandsandmoreresolvedpeakswithrespecttoNIR,theuseofarathersim-ple method for variable selection that works in the same domain ofthe original signal could be suitable. The spectral regions retained

by feature selection algorithms were confirmed as significant foranthocyanins speciationby a deeper NMRinvestigation on selectedsamples, conducted also with 2D experiments.

Finally, 2D correlation analysis was applied to FT-NIR and 1HNMRmatrices in order to recognise the spectral zones bringing thesame chemical information, especially those that were associatedto the presence of anthocyanins.

2. Experimental

2.1. Materials

Adulterated wine samples were prepared starting from five dif-

ferent types of wines: a white wine (Trebbiano dellEmilia), a roswine (Lambrusco Salamino), and three red wines (Sangiovese diRomagna, Lambrusco dellEmilia, Montepulciano).

Allthese natural wines, i.e. wines without anyaddiction, wereused as a basis to prepare two sets of samples by adding two dif-ferent liquids: an aqueous solution of anthocyanins extracted fromblack rice (ZuhHai Golden Land Natural Colors Co., Ltd) having con-centration of 5000ppm,and the wineRossissimo, that is a commonblending wine in Italy.

2.2. Instrumentation

The NIR transmissionspectra were acquired using a Bruker MPA

Multi Purpose FT-NIR Analyzer spectrophotometer, each spectrumresulting from theaverage of 200 scans in the region between 4000and 12500 cm1, with a resolution of 2cm1.

NMRspectra were recorded with a Bruker FT-NMR Avance 400spectrometer (Broad Band5 mm probe, inverse detection). Nominalfrequencies were 400.13MHz for 1H and 100.61MHz for 13C.

2D homonuclear shift correlation (H,H-COSY) spectra wereacquired using the Bruker pulse sequence cosygpqf implying gra-dient pulses for selection with second flip angle being 90. Thephase sensitive 2D 1H,13C HSQC was performed via double INEPTtransfer using Echo/Antiecho-TPPI gradient selection; decouplingduring acquisition was performed using TRIM pulses in INEPTtransfer with multiplicity editing during the selection step [36]. For2D H,X-Hetero Correlated Spectroscopy HMBC proper parameterswere used as suggested by literature [37].


3/13

E. Ferrari et al. / Analytica Chimica Acta 701 (2011) 139151 141

Table 1

Colour Index values reached by the original wines after addition of anthocyanins (a), and results of the chemical characterisation of the corresponding wine samples as foralcoholic content (b) and pH (c). The samples highlighted in grey and marked with codes Nt, Br and Rs werealso submitted to NMRqualitative inspection.

(a)

Wine cultivarNatural CI(natural samples in set R)

Final CI for B-set samples Final CI for R-set samples

Trebbiano0 5 6 7 8 5 6 7 8L. Salamino1 (Nt) 5 6 7 8 5 6 7 8Sangiovese4 5 6 7 8 5 6 7 (Rs) 8L. dellEmilia5 6 7 8 (Br) 6 7 8Montepulciano8

(b)

Alcohol content (%, distillation)Reg.CEE n 2676/90 G.U.CE n L272/3.10.90, all 3 REGCEE n 128/04 G.U.CE n L19/3/27.01.04

12.03 12.07 12.06 12.07 12.06 12.08 12.08 12.07 12.0710.36 10.43 10.46 10.47 10.48 10.55 10.60 10.64 10.7011.85 11.84 11.88 11.85 11.85 11.88 11.87 11.88 11.9010.53 10.58 10.61 10.64 10.59 10.65 10.7212.91

(c)

pH (potentiometry)Reg.CEE n 2676/90 G.U.CE n L272/3.10.90, all 24

3.15 3.22 3.21 3.20 3.19 3.29 3.29 3.28 3.293.60 3.53 3.51 3.47 3.45 3.55 3.53 3.53 3.513.20 3.19 3.18 3.17 3.16 3.21 3.21 3.22 3.223.50 3.51 3.50 3.48 3.50 3.50 3.503.75

2.3. Procedures

2.3.1. Samples preparation procedure

The colour of red wines is commonly defined by a parametercalled Colour Index (or Colour Intensity, CI) defined by Sudraudin 1958 [1,38] which essentially accounts for wine anthocyaninscontent. The CI, which varies widely depending on the origin andage of the wine, is calculated starting from the intensity of absorp-tion of the sample in the visible region of the electromagneticspectrum at =420nm (A420) and =520nm (A520), where theabsorptionmaximaofgreenandbluecolours(representing,respec-tively, the red and yellow components of wine colour) are located.CI is then defined as:

CI = (A420+A520) (number of dilutions) (2.1)

In oenology it is common practice to correct the colour of winesby the addition of blending wine, up to the desired CI value.

Starting from these considerations, the 5 natural wines previ-ously presented have been added with an amount of black ricesolution or Rossissimo to achieve CI values between 5 and 8, in a

manner to be representative of most of the red wines present onthe market.The first set of 15 samples (named B-set) was prepared by

adding to each wine different amounts of black rice solution untilthe desired CI value. The second set of 15 samples (named R-set)was prepared by adding ofRossissimo, up to the desired CI value,as it is common during the corrective procedures in winery. It hasto be underlined that, for classification purposes, we included inthe R-set also the five natural wines without any addiction, sinceall the R-setsamples contain only grapevineanthocyanins. Table 1aincludes the30 added samples together with the 5 natural wines.

Table 1 also reports a basic chemical characterisation of the setof samples. Thetablecellsreporting thevalues foralcoholic content(Table 1b) and pH (Table 1c) match to the corresponding samplesin the analogue position ofTable 1a.

2.3.2. NIR spectra acquisition

The signals of all samples were acquired using two types ofquartz cuvettes with different optical paths: the cuvette type S(short), with an optical path of 0.2 mm, and the cuvette type L(long), with an optical path of 1 mm. This latter is the most fre-quent for similar works in literature [19,39,40]. We have chosento use two different cuvettes starting from considerations raisedby the visual observation of some preliminary spectra. In fact, itwas highlighted that cuvette L, while producing a stronger sig-nal and better defined peaks, presents the problem of bringing thedetector to saturation in the spectral regions around 5200cm1

(corresponding to combination bands of water) and 4050 cm1. Onthe other hand, cuvette S does notgive saturation problems butthecorresponding spectra have relatively low values of absorbance,in particular at about 5600cm1 and in the 750010600cm1

wavenumber range. Hence, we preferred not to establish a prioriwhich cuvette was better for the problem under investigation.

TheNIR background signalwas obtainedwith theemptycuvetteholder. The system was thermostated at 35 C, as suggested byCozzolinoet al. [41], to increase thereproducibility of themeasure-ments. Before spectra acquisition, each wine sample was subjectedto magnetic stirring for 10min to dissolve the CO2 bubbles that

could distort the measurement.For each of the 35 wine samples, four repeated measurements

were acquired for a total of 140 spectra. Each repeated spectrumwas recorded on a different aliquot of sample, which was inde-pendently degasified and then used to fill the cuvette. The order ofexecution of experiments was randomised in order to reduce therisk of introducing systematic variations in the data.

2.3.3. NMR spectra acquisition

As for NMRspectroscopy, each wine sample (0.5 mL) was addedof 0.1mLD2O, placed in a 5 mm NMR tube and directed to anal-ysis. An internal lock on the deuterium was used for all spectra.All experiments were performed at 300K and nonspinning; chem-

ical shifts were referred to TSP.1

H NMRdata were acquired using


4/13


Fig. 1. Composite signal, SL.

the water suppression pulse sequence zgcppr (Bruker Library),

with the following parameters: spectral width 10ppm, 32k datapoints; 0.51 s relaxation delay; collection of 192 transients. TheNMRspectra were processed with X-NMR (Bruker), and they wereFourier transformed with a FT size of 64k and 0.3 line-broadeningfactor. In order to manipulate data as littleas possible,an automaticphase correction was applied to each sample.

For each of the 35 wine samples, two repeated measurementswere acquired for a total of 70 spectra. With a similar procedureused for NIR spectroscopy, each repeated spectrum was recordedon a different aliquot of sample, paying attention to randomise theorder of execution of the experiments.

With the aim to investigate the effect of the addition of antho-cyaninsfromblackriceorgrapevineontheNMRspectrumofawinesample, a specific characterisation by further NMRanalyses were

thus conducted. To this goal, three samples highlighted in grey inTable 1a were selected from thewholeset: sampleNt, constitutedby a natural (not adulterated) wine sample, sample Br, adulteratedwith black rice solution, and sample Rs, added with Rossissimo.

The selected samples were pretreated and acquired as follows:20mLof each sample were concentrated under reduced pressureat room temperature; an amount of 300 mg was then added with1 mLof MeOD, vigorously stirred and centrifuged. Then, 500L ofthe supernatant were placed in a 5 mm NMRtube and directed toanalysis.

2.3.4. Data organisation

The NIR spectra acquired with S and L cuvettes were organisedin two distinct matrices before data analysis.

In addition, an original signal processing was applied on S andL spectra to condense the information obtained with both cuvettesin a single signal. The purpose of this approach was to possi-bly improve the classification results between adulterated andnon-adulterated samples. This signal processing consisted in thebuilding of a composite signal that hopefully presents the advan-tages of both kinds of signals, butnot their defects. In particular, weintended to obtain a signal where the saturated absorbance valuesfromcuvetteLaredeleted,buttherelativelyhighabsorbancevaluesin other spectral regions are maintained. Then, the absorbance val-ueshigher than 2 units foreach spectrum acquired with thecuvetteL were setequalto 2, givingrise to a truncated signal(shown witha grey dotted line in Fig. 1). The corresponding spectrum acquiredwith the cuvette S, represented as a grey dashed line in the figure,

was added to this signal to form the composite signal, called SL,

shown in black. The 140 SL signals obtained were organised in thecorresponding SL matrix.

The matrix of NMRsignals was built too and then it was sub-jected to alignment. This step was necessary because no samplepreparation was done, whereas, pH differences among the winesamples can induce some chemical shift scatter in the spectra. Inthe literature, various algorithms arecommonly used to correct thechemical shift of NMRspectra [30,4244]. In the present case, wepreferred to use the novel algorithm icoshift, that was specificallydesigned by Savorani et al. [45] for solving signal alignment prob-lems in metabonomic NMRdata analysis. The whole spectra werealigned using the average spectrum as a reference, then, the regionbetween 6.5 and 9.5 ppm was maintained for the following chemo-metric analyses. That interval is the so called aromatic region, inwhich are located the chemical shifts of phenolic compounds, inparticular the aglycone moiety of anthocyanins [27,31,32].

2.3.5. Explorative PCA

NIRand NMRspectral matrices were submitted to meancenter-ing before explorative analysis PCA, that was used to identify thepresence of possible outlier samples and/or of objects clusters.

Before applying the classification methods, the NIR spectramatrix was randomlydivided in a training set and in a test set,used

for external validation, paying attention to maintain the replicatemeasurements of each sample in the same set. The training set wascomposed by 21 samples, 9 (4 replicates= 36 spectra) of class Band 12 (4 replicates= 48 spectra) of class R; the test set was com-posed by 14 samples, 6 (4 replicates= 24 spectra) of class B and8 (4 replicates= 32 spectra) of class R. For comparison purposes,the samples of NMRmatrix were divided in training and test setsincluding the same samples respectively in the same set, in thiscase with 2 replicate spectra instead of 4 for each sample.

2.3.6. Classification methods

For the selection of the optimal values of the meta-parameters(number of LVs, variables to select, signal pretreatments, wavelettypes, etc.), for all the classification algorithms described in this

section, we used internal cross-validation (CV). In particular, inthis application, was used a cross-validation with three cancella-tion groups, forcing the algorithm to keep the replicate spectra inthe same group. The resulting classification performances (on thetrainingset,inCV,andonthetestset)wereexpressedaspercentageefficiency (EFF), which is the geometric mean of sensitivity (SENS,the percentage of objects of each class accepted by the class model)and specificity (SPEC, the percentage of objects of the other classescorrectly rejected by the class model) values [46]. The best modelspresented in this work were all selected by their CV efficiency.

For both NIR and NMR spectral matrices, PLS-DA was usedto build the classification models [47]. As for NIR dataset, 16pre-processing methods were applied before classification, sinceit is usually not possible to establish in advance the one work-

ing better [48]: none (N), first derivative (D1), second derivative(D2), linear detrend (Det), quadratic detrend (Det2), smoothing (S),Multiplicative Scatter Correction (MSC), Standard Normal Variate(SNV), meancentering (mncn), D1+ mncn, D2+ mncn, Det + mncn,Det2+ mncn, S + mncn, MSC + mncn,SNV + mncn. For the same rea-son, for NMRdataset 13 pre-processing methods were applied:N, D1, S, MSC, SNV, mncn, D1 + m ncn, S + m ncn, MSC + m ncn,SNV+ mncn, D1+ S +mncn, MSC +S +mncn,SNV+S +mncn.

Since the information related to anthocyanins could be over-whelmed by uninformative variation, such as the chemicalcomposition of thefive starting wines, we alsoappliedclassificationmethods with variable selection.

To this aim, a novel algorithm based on the Fast Wavelet Trans-form (FWT), called WILMA-D, was applied to NIR and NMRspectra

matrices.


5/13


FWT is a decomposition method in the wavelet domain whichrecursively splits the low and the high frequency contents ofthe signal into two orthogonal sub-spaces applying two filters tothe signal, which correspond to a given wavelet, chosen from alibrary of different wavelets. A low-pass filter and an high-pass fil-ter maintain, respectively, the low frequency contributions in theapproximations vector (A1), and the high frequency contributionsin thedetails vector (D1).A1 andD1 vectors are complementary, i.e.the information content of the signal is completely preserved afterthe decomposition, but it is split in two contributions of lengthcorresponding to a half of the original signal. The decompositionkeeps going on at the following level A1 is halved intoA2 and D2by applying the same two filters and so on: at each decomposi-tion level the approximation vectorAn is almost halved intoAn+1and Dn+1. The maximum possible decomposition level is reachedwhen approximation and detail vectors cannot be further halved.

The signal variables in the wavelet domain namely waveletcoefficients, that constitute a set of independent variables areselected in WILMA-D according to best predictive performance,expressed as efficiency, of the derived PLS/MLR regression models,as evaluated by cross-validation. Since WILMA-D is the modifiedversion for discrimination purposes of the calibration algorithmWILMA, the reader is referred to comprehensive references [34,35]

to find a deeper description ofWILMA. In this paper are reportedonly the main operating principles ofWILMA, that are shared withWILMA-D, and the specific conditions that were set. In particular,in WILMA-D the response matrix Y is simply a vector composed ofinteger numbers, i.e. 1, 2, . . ., n, where n is the number of the dif-ferent classes. Each observation has the value corresponding to theclass it belongs to.

Schematically,thealgorithmworksasdescribedinthefollowingprocedure:

the matrix of signals is decomposed by means of FWT until itsmaximum level of decomposition, which is depending on thelength of the signals;

the wavelet coefficients of each decomposition level are ranked

according to a specific parameter, chosen by the user amongthe following ones: variance (V), squared correlation coefficient(R2), squared Spearmans correlation coefficient (S2) or squaredcovariance (C2) with the response variable;

for each level, the optimal number of wavelet coefficients isiteratively selected by a proper cross-validation procedure(Leave-One-Out/Venetian Blind/Random Groups/ContiguousBlocks/Custom) using both MLR and PLS as regression meth-ods. The discriminant version of WILMA was developed tochose the optimal number of wavelet coefficients based on thehighest efficiency, where EFF is calculated by assigning eachobject to the class that is modeled, i.e. rounding up the valuecalculated/predicted to the nearest integer;

the decomposition level giving the best efficiency in cross-

validation is chosen as the best one.

Since it is generally impossible to know in advance the optimalcombination of theWILMA-D parameters, it is appropriate to cycleover different possible combinations in order to find the classifi-cation model having the higher efficiency. For both NIR and NMRspectra datasets,WILMA-DwasappliedwithPLS asregression tech-nique on autoscaled wavelet coefficients. Sixty models for eachdataset were obtained by setting R2, C2 and Vas ranking parame-ters and 20 wavelets, from three families, daubechies (db), symlets(sym) and coiflets (coif): db1db10, sym4sym8, coif1coif5.

Since the NMR spectra show better resolved peaks than NIRspectra where, on the contrary, the information related to dif-ferent chemical species is heavily superimposed for NMRdata

only, WILMA-D with MLRas regression technique was also tested,

using the same sixty combinations of parameters. For the samereason, and considering that the use of a simple method for vari-able selection that works in the same domain of the original signalcould be suitable as well, a further variable selection method wastestedon NMRspectra, i.e. the interval version of PLS-DA (iPLS-DA).iPLS-DA, as implemented in thePLS Toolbox forMatlab [49,50], cal-culates local PLS-DA models on subintervals of equal width of thefull spectrum region, then the predictive performances of all thelocal models, obtained by using both forward and backward pro-cedures, are compared. The combination of intervals that give thebest performance is identified by the lowest RMSECV value.

ForthepresentiPLS-DAanalysisthebestpre-processingmethodas emerged by PLS-DA models was used to pre-process the spec-tral matrix. The aromatic spectral range between 6.5 and 9.5 ppm,comprising 2458 variables, has been split into a different numberof intervals, consisting of 20, 50 or 80 variables.

2.3.7. Correlation analysis

Finally, the correlation analysis between the matrices of the SLcomposite FT-NIR spectra and the 1H NMRspectra was carried outusingtheapproachproposedby Sasic andOzaki [51] andOzaki [52].To this aim, a Matlab function, called twodcorrmap, was properly

developed. The replicate measurements of each matrix were aver-agedtohavethesamenumberofsamples,i.e.35;thematriceswerethen analysed by twodcorrmap without applying any pretreatment.The output is a matrix where each element ij is given by the cor-relation coefficient of the i-esim NIR variable with the j-esim NMRvariable. In the corresponding 2D correlation map, to be clear, allthe elements whose absolute value is lower than a given threshold,Rlim, are set equal to 0. The Rlim depends on the probability levelspecified and on the number of signals. In the present work, thethreshold value was set at 95% probability level.

All calculations were performed in Matlabver. 7.0 (Math-works); WILMA-D was written using some routines from PLSToolbox Ver 5.2 (Eigenvectors) and Wavelet Toolbox ver. 3.0(Math-works).

3. Results and discussion

3.1. Multivariate analysis of NIR spectra

3.1.1. Explorative PCA

Exploratory analysis PCA on meancentered NIR spectra hasshown similar patterns for the spectra acquired with both S(0.2 mm) and L (1 mm) cuvettes. In particular, the scores plot ofthe first two PCs (not shown) shows that the B and R sets are over-lapped in the PCs space. The analysis of the following componentsdoes not show a separation of objects based on the origin of antho-cyanins in the sample too. On the contrary, from the PCA model,wecan partly distinguish the spectra based on the date of acquisition.

The change over time wasascribable to a drift in theemission of theNIR source, but the precaution to randomise the order of executionof the experiments has prevented the occurrence of a confoundingeffect between the addition of the different anthocyanins and thedate of acquisition of the spectra.

Fig.2 showsthescoresplotofPC2vs.PC3forthedatasetacquiredwith cuvette L, wherea cleardifferentiation ofsamples basedon thenatural CI of wine can be seen. This differentiationis theresultof anevident matrix effect. In fact, to simulate the adulteration process,it was decided to add anthocyanins on wines different for naturalCI. Inevitably, these wines are different from each other also forthe presence and concentration of a large number of constituents(i.e. SO2, ethanol, sugars and organic acids, among others), whosecontributionstotheNIRsignalshouldbesuperimposedtotheinfor-

mation related to anthocyanins. This fact shows that the goal of the


6/13


Fig. 2. Scores plot of PC2 vs. PC3 for meancentered NIR spectra. Thedifferent sym-bols stand for thefive original wines, having a differentnatural CI.

work is far from trivial, because it is intended to develop a quickmethod that does notinclude anypreparation of the samplefor theelimination of interfering substances.

3.1.2. Classification by PLS-DA

The spectral datasets from S andL cuvettes were then processedby PLS-DA. Among all the obtained classification models, the bestperforming ones were selected based on their efficiency in CV,thenthemodels were validatedwith theexternaltest sets. The results ofthetwobestmodelsforcuvettesSandLdataarereportedin Table2.Themodelon thecuvette S data required fewer latentvariables; thesame trend was observed for most of the applied pre-processingmethods. However, cuvette L produced more satisfactory results,although not excellent.

The matrix of the 140 SL signals was processed by PLS-DA: thecorresponding results for the best pre-processing method are alsoreported in Table 2. Compared to the results that were obtainedusing the spectra from cuvettes L and S, the classification on SLcomposite signals has generally led to some improvements in theclassification performance, in terms both of Efficiency and of thesingle Sensitivity and Specificity values.

3.1.3. Variable selection/classification by WILMA-D

The development of classification models with variables selec-tionwasprimarilyaimedtoremovethecontributionofthemajorityof the chemical components that compose the five original wines;such matrix effect constitutes an interference for adulterationdetection.

Since SL signals performed better than S and L spectra alone, thevariable selection was applied to the dataset of SL spectra. The lastcolumn ofTable 2 shows the performance of the best, among the10 examined,WILMA-Dmodel,whichwas obtained using a waveletdb1 atthe fifth levelof decomposition.This model allowed toobtainan improvement of the classification, being the test set now clas-sified with an efficiency of about 70%. In this case, with only 212wavelet coefficients, i.e. descriptors, it was possible to obtain betterresults than those obtained by PLS-DA with 8815 variables.

Fig. 3 reports the original NIR signals of class B with the regionsselected as the most informative to discriminate between the twoclasses of samples highlighted in grey colour (a) and the corre-sponding signals reconstructed from the wavelet coefficients keptafter variables selection withWILMA-D (b). A wide region, between

about 124007200 cm1

, has been selected by the algorithm, that

Table 2

Results of the best models obtained on NIR spectra by PLS-DA for the differentcuvettes and by WILMA-D on SL signals.

Classificationmethod

PLS-DA PLS-DA PLS-DA WILMA-D PLS

Cuvette type S L SL SL Pre-processing Det + mncn Det + mncn SNV + mncn None# LVs (# wav.

coefficients)1 4 7 7 (212)

Conditions set Ranking: C

Wavelet: db1Level: 5

Training setSENS

Class R 52.8 88.9 91.7 94.4Class B 68.8 75.0 87.5 89.6

SPECClass R 68.8 75.0 87.5 89.6Class B 52.8 88.9 91.7 94.4

EFFTOT 60.2 81.7 89.6 92.0

CVSENS

Class R 52.8 69.4 86.1 75.0Class B 62.5 64.6 62.5 64.6

SPEC

Class R 62.5 64.6 62.5 64.6Class B 52.8 69.4 86.1 75.0

EFFTOT 57.4 67.0 73.4 69.6

Test setSENS

Class R 25.0 83.3 66.7 66.7Class B 68.8 46.9 59.4 75.0

SPECClass R 68.8 46.9 59.4 75.0Class B 25.0 83.3 66.7 66.7

EFFTOT 41.5 62.5 62.9 70.7

is probably necessary to correct for the baseline shift. Four fur-ther spectral regions have been selected: the first one is a ratherwide portion of wavenumbers between 6540 and 5620cm1, thesecond one includes the 53105095cm1 range, the third one the44454320 cm1 range and the fourth includes the narrow rangebetween 4045 and 4000cm1. The bands in these regions havebeen identified by other authors as being typical of componentsof wine samples. In particular, Cozzolino et al. [19] demonstratedthat theband centred at about 5200 cm1 is dueto thecombinationof stretch and deformation of the OH group in water and ethanol.The twothin absorption bands in the region44454320 cm1 wereassigned to the combination of CH-stretch and CH-deformationof the CH3 and CH2 groups of ethanol. Also the CH-stretch firstovertones of CH3 and CH2 groups were selected, whose absorp-

tion bands are located in the region 65405620 cm

1. Casiraghiet al. [53] studied the anthocyanins content of grape by applying2D correlation spectroscopy to vis/NIR and FT-NIR data. The cor-relation map showed some positive correlations between vis andNIR spectral data. The visible absorption bands at 575450 nm ofanthocyanins are related at 8440 and 6490 cm1 NIR bands, due tofirst and second overtone of CH stretching. Moreover, the 575 nmvis band also corresponds to the first OH overtone (6490 cm1),associated to carbohydrate and sugars molecules, being the antho-cyanins always present in glycosylated form in grapes. As it can benoticed, there is not an unequivocal correspondence between theselected bands in the best WILMA-D model and the bands recog-nised as peculiar for anthocyanins in literature, but it is interestingto observe that the other nine WILMA-D cycles have selected, as

informative for classification aims, the same spectral regions.


7/13


Fig. 3. Original NIRsignals of class B with selected regions highlighted in grey colour(a) andthe corresponding signalsreconstructed after variablesselection with WILMA-D

(b).

A deeper inspection of the classification results for the test setproved that the misclassified samples mostly correspond to a lowor, sometimes, to an intermediate increase of Colour Index that theaddition of anthocyanins produced, i.e. the samples for which theamount of anthocyanins added from Rossissimo or black rice waslow. The model is therefore able to discriminate better the samplesof thetwo classes that have a higherconcentration of anthocyanins,while it seems to be more difficult to classify the samples modifiedto a lesser extent.

3.2. Multivariate analysis of NMR spectra

3.2.1. Explorative PCAThe data structure of the aligned NMR spectra matrix was

explored by means of PCA after meancentering. The scores plot ofthe first 3 PCs (Fig. 4), explaining about the 88% of total variance,shows that thedifferentiation of wine samples accordingto theori-gin of anthocyanins added is not trivial. The R and B sets of spectraare partially overlapped in the space defined by the principal com-ponents. On the other hand, it can be clearly seen that the samplesare differentiated on the basis of their wine of origin to whichthe additions were made each wine being characterised by a dif-ferent natural CI. As for NIR, exploratory analysis showed that thematrix effect is predominant over the compounds present at lowerconcentrations, such as anthocyanins.

3.2.2. Classification by PLS-DAFollowing the same procedure adopted for the set of NIR spec-

tra, the matrix of NMRsignals has been subjected to classificationby PLS-DA. The results of the best model, chosen in CV and cor-responding to the data matrix pre-processed by a sequence ofthree methods i.e. SNV, smoothing (polynomial order: 0, filterwidth: 15 points) and meancentering are reported in Table 3, thefirst data column. The obtained efficiency values in CV and on thetest set are comparable and both higher than 90%, confirming thatNMR spectroscopy coupled with multivariate classification is ableto recognise the different qualitative nature of anthocyanins fromgrapevineor black rice. Forthe present model,alsothe VIP(VariableImportance in Projection) scores plot is presented (Fig. 5). The VIPscores estimate the importance of each variable in the projection

used in a PLS model; the greater than one rule is generally used to

determine whether a certain variable is actually significant; in thismanner is possible to identify the portions of the spectrum thatare most useful to classification [54]. Fig. 5 shows that the spec-tral regions that are responsible for PLS-DA classification, i.e. thoseabove the threshold value of 1, are mainly located at low chem-

Table 3

Results of thebestmodels obtainedon NMRspectra by usingdifferentclassificationmethods.

Classificationmethod

PLS-DA iPLS-DA WILMA-DMLR

WILMA-D PLS

Pre-processing SNV + S +mncn SNV+S+mncn None None

# LVs (#wav.coefficients)

5 5 (3) 6 (87)

Conditions set Forward50 vars

Ranking: RWavelet: db7Level:4

Ranking: CWavelet: coif3Level:6

Training setSENS

Class R 100 100 100 94.4Class B 100 100 100 100

SPECClass R 100 100 100 100Class B 100 100 100 94.4

EFFTOT 100 100 100 97.2

CVSENSClass R 100 100 100 94.4Class B 83.3 100 100 95.8

SPECClass R 83.3 100 100 95.8Class B 100 100 100 94.4

EFFTOT 91.3 100 100 95.1

Test setSENS

Class R 100 83.3 91.7 91.7Class B 87.5 93.8 87.5 100

SPECClass R 87.5 93.8 87.5 100Class B 100 83.3 91.7 91.7

EFFTOT 93.5 88.4 89.6 95.7


8/13


Fig. 4. Scores plot of the first 3 PCs for meancenteredNMRspectra.

ical shifts in the aromatic region. In particular, almost the entireportion of the signal between 6.5 and 6.97ppm has a significantrole in the classification, as well as the regions centred around7.107.12, 7.187.22, 7.427.46ppm and minor peaks at 8.48, 8.63and 9.40 ppm.

3.2.3. Variable selection/classification by iPLS-DA

For NMR spectra, the iPLS-DA classification with variableselection was investigated. Interval PLS-DA was applied to the

normalised, smoothed and meancentered spectral matrix afteralignment, i.e. using the best pre-processing method as emergedby PLS-DA models. Among the six calculated models obtained on20,50, 80 variables intervals with forward or backward procedures the one giving the best EFF in CV was selected; its performancesarereportedintheseconddatacolumnofTable3. Eveniftheperfor-

Fig. 5. VIP scores for PLS-DA classification model obtained on NMRspectra. The

horizontal dashed line indicates the threshold value.

mances in calibration and in CV of that model reached the 100% ofefficiency, the prediction on the test set is surprisingly poorer thanthatobtainedwithoutanyselectionofvariables.Theincludedinter-valsfor modelcalculationwerethe following:6.506.56, 7.057.11,7.427.48, 8.768.82 ppm.

3.2.4. Variable selection/classification by WILMA-D

Concerning the WILMA-D models, the last two data columnsofTable 3 report the performances of the best models developed

using both MLRand PLS as regression tools. The most satisfactoryperformances on the test set were obtained using PLS as regres-sion method, in a model with six latent variables. The 87 selectedwavelet coefficients led to an efficiency in classification over 95%.However, regardless of the effectiveness of individual models herepresented,NMRspectroscopy generally seemsmuch more efficientthan NIR to identify thepresence, in wine samples, of anthocyaninsthat are not derived from grapevine.

In Fig.6a isreportedtheFWTdecompositionfor thebestWILMA-D PLS model. For the approximation block (A) at the sixth leveland the details blocks (D) of the first six levels, the regions whereare located the coefficients responsible for classification are high-lighted in black. In all plots are represented both classes of signals,but only in A6 plot they differ significantly. For each different vec-

tor the algorithm has selected some precise features; only at thesixthlevel of decomposition, almost the entire approximationsanddetails vectors have been selected. As the level of decomposition islowered, it is noted that the areas selected from the details vectorsbecome more specific and limited. At different stages of decompo-sition, especially at the first three levels, the same spectral regionsidentified by the otherclassification methods were selected. There-fore WILMA-D PLS converges at least in part to the same NMR regions for the optimal classification identified by PLS-DA,iPLS-DA and also by WILMA-D MLR model, which selected threecoefficients located at 6.926.98, 7.377.52 and 9.219.23 ppm ofchemical shift.

Evidently, WILMA-D PLS produces better results than the othervariable selection methods since, although these methods attempt

to select regions that cover the entire spectral range, are not able


9/13


Fig. 6. Results for the best WILMA-D PLS model: (a) approximations and details vectors in the WT domain, where black points highlight the regions where are located thecoefficients responsible for classification; (b) reconstructed average signal foreach class, grey for B and black forR class.

to consider the contributions that account for the overall shape ofthe signal. On the contrary, the wavelet coefficients derived fromWILMA-D PLS model are distributed at various levels throughoutthe signal, thus they detect the fingerprint of the sample. In thissense, the classical PLS-DA, which does not require a selectionof variables, worked in a comparable way. This is confirmed byFig.6b, whichshowstheaveragereconstructedsignalforeachclass.Basically, the signals reconstructed in the original domain usingonly theselectedvariables arevery similar to theoriginalones withregard to their shape, although some peaks, that are not significantfor classification, have disappeared and the profile of the spectralook quite smoothed.

3.3. Qualitative NMR analysis of anthocyanins

A qualitative analysis of1H NMRspectra evidences, for all sam-ples, the occurrence of main signals in the sugar-like (5.53ppm)and in the aliphatic (03ppm) regions. Although anthocyaninsrepresent minor compounds, they are anyway of major interestfor wine classification. As reviewed by Escribano-Bailn et al. [6]the principal anthocyanins extracted from red and black rice arecyanidin-3-glucoside and peonidin-3-glucoside, while malvidin-3-glucoside is usually the predominant anthocyanidin in mostred grapes and consequently in wines, followed by delphinidin-3-glucoside and petunidin-3-glucoside [5557]. The significant


10/13


Fig. 7. 1H NMRspectra of pre-concentrated (a) and plain (b) Nt sample respectively in D2O (17%, v/v) and MeOD (100%) at 300K. At the top right side of the figure thechemical structureand thenumbering schemeof anthocyanins typical ofVitis vinifera are also reported.

regions derived from chemometric analysis match with NMRassignment of anthocyanins, especially the area 6.56.97 is typ-ical of H6 and H8 in many anthocyanin-derivatives; 7.057.11is characteristic of H5 in cyanidin and peonidin; 7.187.22 and7.427.46 can be reasonably attributed respectively to H6 andH8 in malvidin adducts. Peaks around 9 ppm are due to H4[25,58,59].

A semi-quantitative analysis was finally performed on the threeselected samples mentioned in Section 2: Nt (natural, i.e. not adul-terated wine sample), Br (adulterated with black rice solution) and

Rs (added with Rossissimo). The pre-concentration of each sampleandits dissolution in MeOD accomplishesto incrementthe contentin anthocyanins, leading to higher resolution and S/N ratio in thearomatic region. A comparison between the pre-concentrated andthe plain samples of the same wine (Fig. 7), points up a similarcomposition, confirming that the pre-analysis treatment does notdecompose, nor affect the analytes in the matrix. The signal shiftsare due to the change in the solvent system: H2O/D2O vs. MeOD.The addition of black rice solution in sample Br results in the neatdecrease in integrated areas of peaks under the regions A, B, C andD defined in Fig. 8, all reasonably ascribed to aromatic protons inanthocyanin structures, particularly the absence of the singlet cen-tredinregionC(7.6ppm) corresponding to H2 of delphinidin andpetunidin,may be attributed to thedeficiencyof these components

in the extracts from black rice. Even though 1D

1

HNMRspectra are

extremely informative on qualitative composition of the matrices,we have also performed 2D homo-correlation experiments, whichmay provide additional information on minor compounds. Giventhat the COSY cross-peaks do not represent all species present inthe matrix actually there are protons only weakly coupled ornot coupled at all the comparison of the three spectra (Fig. 9)points out that only few correlations are separated. They anywaycorrespond to spin-systems of trace compounds characterised bystrong couplings, i.e. trans double bonds. The correlation peaksat (7.73; 7.10ppm) and (7.20; 6.60ppm) present only in sample

Br are reasonably due to ortho protons in aromatic systems ofanthocyanin-derivatives or adducts. 2D hetero-correlation spectra(HSQC and HMBC) are instead extremely time consuming andlittle informative: only the strongest aromatic 1JCH and 3JCH areobserved consequently with a complete overlie of spectra of thethree samples, independently on the origin of added anthocyanin(spectra not shown).

3.4. Correlation analysis

The correlation analysis for the matrices of composite NIR sig-nals and 1H NMR signals furnished the corresponding 2D map;the only portion in which are present some significant (P=95%,Rlim = 0.33384) correlation peaks is reported in Fig. 10. As for

1H

NMR signals, the peaks showing a significant correlation with NIR


11/13


Fig. 8. 1H NMR stacked plots of pre-concentrated Nt, Br and Rs in MeOD at 300K. Squares (A, B, C and D) point out the main differences between the samples added withblack rice solution with respect to Rossissimo.

peaks are those around 6.92 and 7.11ppm. These peaks can berelated to anthocyanin-derived compounds: actually H2 /H6 pro-tons in hemiacetal forms of malvidin-3-glucoside typically give1H NMR peaks around 6.90ppm [59], and in the same chem-ical shift interval also H8 of malvidin adducts may occur [60].H6 protons of malvidin adducts [55] and peonidin-3-glucoside[61] provide singlets in the chemical shift range 7.107.13ppm.These literature attributions are coherent with the regions thatwere selected by blind classification analysis of 1H NMR spec-tra.

On the contrary, the results from NIR blind analysis were not asmuch satisfactory and, particularly, the attribution of the selectedbands was not straightforward. As a consequence, the 2D correla-tion map could be useful in order to recognise, among the severelyoverlapped NIR bands, the ones underlying the anthocyanins-related information. As for the NIR spectra, the peaks havingpositive correlation with 1H NMR peaks are those around 4200,5280, 5400cm1 and between 7200 and 6800cm1. In particular,the two peaks at 4200 and 5280cm1 were also located in spectralregions selected by the presented WILMA-D model. Moreover, a

Fig. 9. H,H COSY NMRspectra of Nt (a), Br (b) and Rs (c) samples in MeOD at 300K. Triangles, squares and circles highlight thecharacteristic correlations.


12/13


Fig. 10. Portion of the 2D correlation map obtained forcomposite NIR spectra (12,5004000cm1) and 1H NMRspectra (6.59.5 ppm).

negative correlation over 95% with 1H NMRwas found for the NIRpeaks at 44454320 cm1, due to the combination of CH-stretchand CH-deformation of the CH3 and CH2 groups of ethanol.

4. Conclusions

In this work is proposed a novel application of FT-NIR and 1HNMR-basedblindanalyses,aimedtodiscriminatewinesaddedwiththe blending wine Rossissimo from wines adulterated with antho-cyanins extracted from black rice to increase their Colour Index.

The results achieved by NIR spectroscopy showed that arelationship between near infrared spectra and anthocyanins com-

position of wine does exist, even if the classification did not reachexcellent results. In particular, the performance of the classifi-cation obtained seems to be not independent on the amount ofanthocyanins added to the samples. It is indeed known that NIRspectroscopy suffers from scarce sensitivity in determining com-pounds present at low concentrations. A further problem, not sounexpected, has been arisen from the fact that it was decided toadd anthocyanins on different original wines to simulate the realprocess of adulteration, resultingin an evident matrix effect, whosecontribution to the NIR signal was probably overwhelming.

On thecontrary, 1HNMRspectroscopy, although certainly beingless cheap and fast than NIR, was successfully applied on unex-tracted samples for the same goal. This fact shows that the entireNMR spectrum can be analysed as a fingerprint to investigate the

chemical natureof a sample andto predict features that areusually

obtained with time consuming laboratory tests, with considerableadvantages of speed and ease of acquisition of experimental data.

The most satisfactoryresults were obtained with the model cal-culated using the algorithm WILMA-D, which carries out beforethe classification the variables selection in the wavelet domain.This method yielded a classification efficiency greater than 95%in validation. Moreover, by using WILMA-D as variable selectionalgorithm, no previous data pre-processing is necessary and veryparsimonious models are generally obtained.

The aromatic region of three particular samples was furtherinvestigatedfrom the qualitative point of view by both 1D 1H NMR,which is extremely informative on qualitative composition of the

analytical matrix, and by 2D homo- and hetero-correlation NMRexperiments, which may be useful to provide additional informa-tion on minor compounds. The qualitative investigation markedsome differences in the spectra belonging to samples adulteratedwith anthocyanins of different origin in the same spectral regionsretained by the feature selection algorithms.

Finally, the 2D correlation analysis between FT-NIR and 1HNMR spectra highlighted some significant correlations, allow-ing the use of the more sensitive and intelligible NMR analysisfor the interpretation of overlapped NIR bands. Moreover, theanthocyanins-related peaks in the NMRspectrum resulted to havea correspondence in some NIRregions,whichwere included amongthose selected by the previously presented classification algo-rithms. Interestingly, these regions are not yet reported in the

literature as being related to anthocyanic compounds.


13/13


Acknowledgements

The authors wish to thank Dr. Ilaria Gibertini for providingsignificant assistance during the experimental phase. We arealso thankful to Centro Interdipartimentale Grandi Strumenti(CIGS) of the University of Modena and Reggio Emilia and to theFondazione Cassa di Risparmio di Modena which supplied NMRspectrometer.

References

[1] J.S. Jensen, S. Demiray, M. Egebo, A.S. Meyer, J. Agric. Food Chem. 56 (2008)11051115.

[2] I.J. Kosir, B. Lapornik, S. Andrensek, A.G. Wondra, U. Vrhovsek, J. Kidric, Anal.Chim. Acta 513 (2004) 277282.

[3] P. Ribereau-Gajon, Y. Glories, A. Maujean, D. Dubordieu,Trattatodi enologia II,Edagricole, Bologna, 2007.

[4] EEC regulation 1493/99.[5] S.N. Ryu, S.Z. Park, C.T. Ho, J. Food Drug Anal.6 (1998) 729736.[6] M.T. Escribano-Bailn, C. Santos-Buelga, J.C. Rivas-Gonzalo, J. Chromatogr. A

1054 (2004) 129141.[7] E.S.M. Abdel-Aal , J .C. Young, I. Rabalski , J. Agric. Food Chem. 54 (2006)

46964704.[8] T. Laokuldilok, C.F.Shoemaker, S. Jongkaewwattana, V. Tulyathan,J. Agric.Food

Chem. 59 (2011) 193199.[9] A. Versari, R.B. Boulton, G.P. Parpiniello,Food Chem. 106 (2008) 397402.

[10] A.B. Bautista-Ortn, J.I. Fernndez-Fernndez, J.M. Lpez-Roca, E. Gmez-Plaza,

J. Food Comp. Anal. 20 (2007) 546552.[11] E. Revilla, E. Garca-Beneytez, F. Cabello, G. Martn-Ortega, J.M. Ryan, J. Chro-

matogr. A 915 (2001) 5360.[12] J.H. Lee, N.S. Kang, S.O. Shin, S.H. Shin, S.G. Lim, D.Y. Suh, I.Y. Baek, K.Y. Park, T.J.

Ha, Food Chem. 112 (2009) 226231.[13] C.R. Welcha, Q. Wub, J.E. Simon, Curr. Anal. Chem. 4 (2008) 75101.[14] F.Qiu, J.Luo, S.Yao, L.Ma, L. Kong, J. Sep. Sci. 32 (2009) 21462151.[15] S. Hillebrand, E.C.Montilla, N. Khler, P. Winterhalter, Agro-Food Ind. Hi-Tech

20 (2009) 5255.[16] F.F. Bauer, J. Kossmann, K.R. Koch, K.H. Esbensen, Anal. Chem. (March) (2008)

13711379.[17] L.J. Janik, D. Cozzolino, R. Dambergs, W. Cynkar, M. Gishen, Anal. Chim. Acta

594 (2007) 107118.[18] V. Di Egidio, N. Sinelli, G. Giovanelli, A. Moles,E. Casiraghi,Eur. Food Res. Tech-

nol. 230 (2010) 947955.[19] D. Cozzolino, M.J. Kwiatkowski, M. Parker, W.U. Cynkar, R.G. Dambergs, M.

Gishen,M.J. Herderich, Anal. Chim. Acta 513 (2004) 7380.[20] D. Cozzolino, W.U. Cynkar, R.G. Dambergs, M.D. Mercurio, P.A. Smith, J. Agric.

Food Chem. 56 (2008) 76317636.[21] N. Sinelli, A. Spinardi, V. Di Egidio, I. Magnani, E. Casiraghi, Postharvest Biol.

Technol. 50 (2008) 3136.[22] G. Zsivanovits, D. Ludneva, A. Iliev, AIP Conf. Proc. 1203 (2010) 556561.[23] M. Blanco, M. Castello, A. Peinado, R. Beneyto, Anal. Chim. Acta 581 (2007)

318323.[24] C.W.Huck,W. Guggenbichler, G.K.Bonn, Anal.Chim. Acta 538 (2005)195203.[25] I.J. Kosir, J. Kidric, Anal. Chim. Acta 458 (2002) 7784.[26] N. Ogrinc, I.J. Kosir,J.E. Spangenberg, J. Kidric, Anal. Bioanal. Chem. 376 (2003)

424430.[27] M. Anastasiadi, A. Zira, P. Magiatis, S.A. Haroutounian, A.L. Skaltsounis, E.

Mikros, J. Agric. Food Chem. 57 (2009) 1106711074.

[28] M.A. Brescia, V. Caldarola, A. De Giglio,D. Benedetti, F.P. Panizzi, A. Sacco,Anal.Chim. Acta 458 (2002) 177186.

[29] Y.Y. Du, G.Y. Bai, X.Zhang, M.L. Liu, Chin. J. Chem. 25 (2007) 930936.[30] F.H. Larsen, F. van den Berg, S.B. Engelsen, J. Chemom. 20 (2006) 198208.[31] G.E.Pereira, J.P. Gaudillere,C. VanLeeuwen, G.Hilbert,O. Lavialle,M. Maucourt,

C. Deborde, A. Moing, D. Rolin, J. Agric. Food Chem. 53 (2005) 63826389.[32] G.E.Pereira,J.P.Gaudillere, C. VanLeeuwen, G.Hilbert,M. Maucourt,C.Deborde,

A. Moing, D. Rolin, Anal. Chim. Acta 563 (2006) 346352.[33] E.F. Boffo, L.A. Tavares, M.M.C. Ferreira, A.G. Ferreira, LWT-Food Sci. Technol.

42 (2009) 14551460.[34] M. Cocchi, R. Seeber, A. Ulrici, J. Chemom. 17 (2003) 512517.

[35] A.Ulrici, M.Cocchi,G. Foca,C. Durante,A. Marchetti, L.Tassi,in:M.P.Colombini,L. Tassi (Eds.), New Trends in Analytical, Environmental and Cultural HeritageChemistry, Research Signpost, Trivandrum, India, 2008, pp. 77136.

[36] W.Wilker, D. Leibfritz,R. Kersebaum,W. Bermel, Magn.Reson.Chem. 31 (1993)287292.

[37] A. Bax, M.F. Summers, J. Am. Chem. Soc. 108 (1986) 20932094.[38] P. Sudraud, Ann. Technol. Agric. 7 (1958) 203208.[39] D. Cozzolino, H.E. Smith, M. Gishen, Anal. Chim. Acta 51 (2003) 77037708.[40] L. Liu, D.Cozzolino,W.U.Cynkar,R.G.Dambergs,L. Janik,B.K.ONeill,C.B.Colby,

M. Gishen,Food Chem. 106 (2008) 781786.[41] D. Cozzolino, L. Liu,W.U. Cynkar, R.G. Dambergs,L. Janik,C.B. Colby,M. Gishen,

Anal. Chim. Acta 588 (2007) 224230.[42] F.J. Wyzgoski, L. Paudel, P.L. Rinaldi, R.N. Reese, M. Ozgen, A.Z. Tulio Jr., A.R.

Miller, J.C. Scheerens, J.K. Hardy, J. Agric. Food Chem. 58 (2010) 34073414.[43] S. Rochfort, V. Ezernieks, S.E.P. Bastian, M.O. Downey, Food Chem. 121 (2010)

12961304.[44] S.Hong-Seok,H. Geum-Sook, M.K. Ki,A. Hyuk-Jin,P. Won-Mok,F. VanDenBerg,

H. Young-Shick, L. Cherl-Ho, J. Agric. Food Chem. 57 (2009) 14811490.[45] F. Sav orani, G. Tomasi, S.B . Eng elsen, J. Mag n. Reson. 2 02 (2010) 190

202.[46] M. Forina, P. Oliveri, H. Jger, U. Rmish, J. Smeyers-Verbeke, Chemom. Intell.

Lab.Syst. 99 (2009) 127137.[47] L.A. Berrueta, R.M. Alonso-Salces, K. Hberger, J. Chromatogr. A 1158 (2007)

196214.[48] . Rinnan, F. van den Berg, S .B. Engelsen, Trends Anal. Chem. 28 (2009)

12011222.[49] L. Nrgaard, A. Saudland, J. Wagner, J.P. Nielsen, L. Munck, S.B. Engelsen, Appl.

Spectrosc. 54 (2000) 413419.[50] R. Leardi, L. Nrgaard, J. Chemom. 18 (2004) 486497.[51] S . Sasic, Y. Ozaki, Anal. Chem. 73 (2001) 22942301.[52] Y. Ozaki, in: H.W. Siesler, Y. Ozaki, S. Kawata, H.M. Heise (Eds.), Near-Infrared

SpectroscopyPrinciples, Instruments, Applications, Wiley-VCH, Weinheim,Germany, 2002, pp. 163178.

[53] E. Casiraghi, L. Bodria, R. Guidetti, N. Sinelli, R. Beghi, G. Cabassi, Ital. FoodBeverage Technol. 50 (2007) 510.

[54] I.G. Chong, C.H. Jun, Chemom. Intell. Lab. Syst. 78 (2005) 103.[55] F. He, L.Mu, G.-L. Yan, N.-N. Liang, Q.-H. Pan, J. Wang, M.J. Reeves, C.-Q. Duan,

Molecules 15 (2010) 90579091.[56] I. Ginjom, B. DArcy, N. Caffin, M. Gidley, Food Chem. 125 (2011) 823834.[57] A.B.Cerezo,E. Cuevas,P. Winterhalter,M.C. Garcia-Parrilla,A.M. Troncoso,Food

Res. Int. 43 (2010) 15771584.[58] N. Mateus, A.M.S. Silva, J. Vercauteren, V. de Freitas, J. Agric. Food Chem. 49

(2001) 48364840.[59] N. Mateus, E. Carvalho, A.R.F. Carvalho, A. Melo, A.M. Gonzlez-Params, C.

Santos-Buelga, A.M.S. Silva, V. De Freitas, J . Agric. Food Chem. 51 (2003)277282.

[60] N. Mateus, A.M.S. Silva, C. Santos-Buelga, J.C. Rivas-Gonzalo, V. de Freitas, J.Agric. Food Chem. 50 (2002) 21102116.

[61] A.Kirca, M. Ozkan, B. Cemeroglu, Food Chem. 101 (2007) 212218.

NIR Red Wines

Documents