Computers and Electronics in Agriculture 94 (2013) 1–11

Contents lists available at SciVerse ScienceDirect

Computers and Electronics in Agriculture

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

Determination of the botanical origin of honey by sensor fusion ofimpedance e-tongue and optical spectroscopy

0168-1699/$ - see front matter � 2013 Elsevier B.V. All rights reserved.http://dx.doi.org/10.1016/j.compag.2013.03.001

⇑ Corresponding author. Tel.: +351 289800900; fax: +351 289800066.E-mail address: rguerra@ualg.pt (R. Guerra).

1 Current address: Department of Food Science & Technology, Oregon StateUniversity, Corvallis, OR, United States.

Pablo A. Ulloa a,1, Rui Guerra b,⇑, Ana M. Cavaco b, Ana M. Rosa da Costa c, Ana C. Figueira a,Amadeu F. Brigas c

a CIEO – Centro de Investigação sobre o Espaço e Organizações, Universidade do Algarve, Campus de Gambelas, 8005-326 Faro, Portugalb CEOT – Centro de Electrónica, Optoelectrónica e Telecomunicações, Faculdade de Ciências e Tecnologia, Universidade do Algarve, Campus de Gambelas, 8005-139 Faro, Portugalc CIQA – Centro de Investigação em Química do Algarve, Departamento de Química e Farmácia, Faculdade de Ciências e Tecnologia, Universidade do Algarve, Campus de Gambelas,8005-139 Faro, Portugal

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

Article history:Received 28 August 2012Received in revised form 22 January 2013Accepted 6 March 2013

Keywords:Botanical originHoneyChemometricsElectronic tongueSensor fusionSpectroscopy

The aim of this study was to discriminate four commercial brands of Portuguese honeys according totheir botanical origin by sensor fusion of impedance electronic tongue (e-tongue) and optical spectros-copy (UV–Vis–NIR) assisted by Principal Component Analysis (PCA) and Cluster Analysis (CA). We havealso introduced a new technique for variable selection through one-dimensional clustering which provedvery useful for data fusion. The results were referenced against standard sample identification by classicalmelissopalynology analysis. Individual analysis of each technique showed that the e-tongue clearly out-performed the optical techniques. The electronic and optical spectra were fitted to analytical models andthe model coefficients were used as new variables for PCA and CA. This approach has improved honeyclassification by the e-tongue but not by the optical methods. Data from the three techniques was thenconsidered simultaneously. Simple concatenation of all matrices did not improve the classificationresults. Multi-way PCA (MPCA) proved to be a good option for data fusion yielding 100% classificationsuccess. Finally, a variable selection method based on one-dimensional clustering was used to definetwo new approaches to sensor fusion, and both yielded sample clusters even better defined than usingMPCA. In this work we demonstrate for the first time the feasibility of sensor fusion of electronic andoptical spectroscopy data and propose a new variable selection method that improved significantly theclassification of the samples through multivariate statistical analysis.

� 2013 Elsevier B.V. All rights reserved.

1. Introduction

The determination of the geographical and botanical honey ori-gin is of the outmost importance not only because of specific leg-islation (Codex Alimentarius Commission, 2001; European Union,2001) but also because of market demands. Additionally, unifloralhoney types have higher market value due to their limited produc-tion and availability.

Melissopalynological analysis, i.e., the identification and quanti-fication of pollen grains in the honey sediment, is the referencemethod when determining the botanical origin of honey. However,this method involves a laborious counting procedure requiringspecialized knowledge and expertise in the interpretation ofresults making the identification of the honeys botanical origin

rather difficult and very time-consuming (Bianchi et al., 2005;Kaškoniene and Venskutonis, 2010).

Hence, there is a need for detection methods that are non-inva-sive, faster, more manageable and adequate for wide-scale samplescreening. In the past decade, electronic devices such as the elec-tronic-tongue (e-tongue) or the electronic-nose (e-nose) have seenincreased use, and coupled with chemometrics have been shown tobe useful tools for the discrimination of honeys botanical origin[for a review see Cavaco et al. (2012)]. E-tongue systems areemerging as fast and easy to-handle measurement techniques withvery promising applications in food process control and qualityevaluation including the determination of honeys floral origin(Dias et al., 2008; Hruskar et al., 2008; Wei et al., 2009; Escricheet al., 2012).

Visible–Near infrared (Vis–NIR) and ultraviolet–visible(UV–Vis) spectroscopies are also widely used for honey screening.They are not particularly sensitive techniques, but they can be veryuseful in probing bulk material with little or no sample preparationas required for real-time measurements. Vis–NIR spectroscopy canbe used in all stages of food processing, from raw material analysis

2 P.A. Ulloa et al. / Computers and Electronics in Agriculture 94 (2013) 1–11

to finished product verification (Nicolai et al., 2007), while UV–Visspectroscopy is useful to detect a wide number of antioxidants,such as phenolic compounds, responsible for the honey biologicalproperties (Antolovich et al., 2000; Estevinho et al., 2008) and con-sidered to be good markers for the authenticity and botanical ori-gin of honey (Tomás-Barberán et al., 2001).

The combined use of two or more techniques should improvesample classification; however, the main challenge is how to com-bine data coming from each technique. There are few examples ofsensor data fusion in food science. Ozer et al. (1995) used sensorsfor the determination of several fruit quality parameters and fusedthe data to improve the efficiency of automated fruit sorting.Li et al. (2007) used the fusion of volatile profiles obtained frome-nose and zNose™ to improve the detection of damaged apples.The fusion of electronic taste and aroma sensing was applied suc-cessfully to the determination of the floral origin of six varieties ofhoney (Hallis et al., 2010). A combination of e-nose and e-tonguemeasurements was used to determine sugar-adulterated honeyamong 18 samples (Zakaria et al., 2011). In all cases the enormousamount of data generated by these systems must be processedusing appropriate multivariate analysis techniques (Cavaco et al.,2012).

The aim of this study was to group Portuguese honey of differ-ent botanical origin using a non-invasive e-tongue and spectros-copy (Vis–NIR and UV–Vis) sensor fusion approach combinedwith chemometrics. This study presents two innovativeapproaches, namely the sensor fusion of electrical and optical mea-surements, and also a new data fusion technique based on a crite-rion for individual variable selection, which shows great potentialfor the classification of honey samples.

2. Materials and methods

2.1. Honey samples

Thirteen commercial Portuguese honey nectar samples of thesame production lot were purchased locally. The samples repre-sented four unifloral types, four producers, four harvest yearsand three geographic origins as stated on the product label. Thefour unifloral honey types were Citrus spp. (orange blossom, OB),Helianthus annuus (sunflower, SF), Lavandula stoechas (French lav-ender, FL) and Arbutus unedo (strawberry-tree, AU). The 13 sampleswere distributed as follows: 3 OB, 4 AU, 3 FL and 3 SF. All sampleswere kept in their original package and left at room temperatureuntil further analysis. All determinations here reported were con-ducted in duplicate or triplicate, the value used for data analysiswas the corresponding mean value.

2.2. Melissopalynological analysis

All honey samples were tested by melissopalynology to confirmthe botanical origin stated on the product label (Lieux, 1980;Louveaux et al., 1978; Von Der Ohe et al., 2004). A random sample(20 g) was dissolved in 40 mL of distilled water at 40 �C, stirred andthen centrifuged for 15 min at 4640g (Kubata, model KN-70, Ja-pan). The supernatant was removed and a mixture (10 mL) of sul-phuric acid and acetic anhydride (1:9) was added to the remainingsediment in each tube. The reaction mixture was placed in a 70 �Cwater-bath for 10 min (acetolysis). After 10 min centrifugation at4176g, the supernatant was removed and a drop of liquefied Kai-ser’s glycerol gelatin (Merck, Darmstadt, Germany) was added tothe pollen grains sediment. Finally, the entire mixture was spreadon a glass slide over an area of about 20 � 20 mm. For each sample/triplicate, at least 800 pollen grains were counted. The pollen

grains were grouped by pollinic types according to Valdés et al.(1987) and the corresponding data was expressed as percentages.

2.3. General sample preparation for e-tongue and optical analysis

Prior to instrumental analysis, honey samples were incubated ina water bath at 50 �C overnight to dissolve any crystalline material,manually stirred to ensure homogeneity, and adjusted with dis-tilled water to a standard total solids content (70 �Brix). This isessentially a standardization procedure to level all the samples tothe same soluble solids content and water content, allowing forconsistent comparisons. Secondary dilutions were further per-formed to optimize the response of each instrument. In each casethe dilution factor was the same for all the samples, and thus thepremises of standardization were not affected. After the dilutionsthe honey samples (�15 mL) were placed in Falcon tubes and al-lowed to equilibrate at room temperature (23 ± 2 �C) for 30 minbefore being used.

2.4. E-tongue measurements

The e-tongue consisted of four working electrodes [10 � 5 mm2

plates of aluminum (Al), gold (Au), platinum (Pt) and indium thinoxide (ITO)]. These were embedded in a composite material andplaced around a gold reference electrode. The electrodes were con-nected to an Agilent 4284A Precision LCR Meter (Santa Clara, CA,USA). The e-tongue was immersed in a 30 mL aqueous solutionprepared by diluting a 20 g honey sample in 20 mL deionizedwater. Each electrode was set to a 50 mV voltage and a20–2000 Hz frequency scan. The data collected at room tempera-ture (23 ± 2 �C) included capacitance (C,nF), conductance (G,lS)and the ratio G/log(x) where x is the angular frequency. Prior toeach measurement, the sensor was conditioned in the honey solu-tion until stable readings at the 20 Hz scanning frequency were ob-served. The sensors were rinsed with deionized water after everymeasurement. The e-tongue analysis generated a total of 252(=3 � 4 � 21) experimental values per sample, i.e., three measure-ments (C, G, and G/log(x)) with four electrodes (Al, Pt, Au and ITO)and 21 frequencies.

Our e-tongue performs impedance measurements, not the moreusual voltammetric or potentiometric measurements. It deliverseffective values for capacitance and conductance, accounting forthe bulk honey and surface effects at interface honey/electrode.Also, the LCR meter is a two port instrument and our setup didnot include a counter-electrode.

2.5. UV–Vis spectroscopy

An aliquot (5 lL) of honey sample diluted in deionized water(100 mL) was placed in a 1 cm path-length quartz cell (UV–VisibleSpectrophotometer, Cary 50 Bio, Varian, Australia) using distilledwater as a blank. In the 200–400 nm wavelength range, a total of201 data points with an average separation of about 1 nm werecollected in triplicate for each honey type.

2.6. Vis–NIR Spectroscopy

Honey samples were placed in 1 mL plastic 1-cm cuvettes. Spec-tra were recorded as absorbance (USB4000-Vis–NIR Spectrometer,Ocean Optics, USA). The first and last points showed a pronouncedinstrumental oscillation and were removed from the analysis lim-iting the data to the 403.9–1032.2 nm range yielding a total of3347 data points. All Vis/NIRS measurements were conducted intriplicates against a blank of distilled water.

P.A. Ulloa et al. / Computers and Electronics in Agriculture 94 (2013) 1–11 3

2.7. Data analysis

The Principal Component Analysis (PCA) and Cluster Analysis(CA) of the preprocessed data (derivation, curve fitting, filtering)were performed with the PLS toolbox (Eigenvector Research,USA) running under Matlab 7.9.0 (The Mathworks Inc., USA). Allother mathematical data transformations were performed in Mat-lab through custom-written functions. The datasets were analyzedfirst in the usual way of performing PCA on the entire spectra (pre-processed or not). The second way involved fitting the spectra tospecific analytical expressions and running the PCA analysis onthe fitting parameters. All matrices were autoscaled after prepro-cessing except in the Multi-way PCA (MPCA) as described below.The agglomerative method chosen for cluster analysis (CA) wasthe K-Nearest-Neighbor (KNN).

2.7.1. Fitting procedure for the e-tongue dataA commonly used model for the capacitance of biological media

is:

CðxÞ ¼ C1 þCs � C1

1þ ðxsÞ2; ð1Þ

where x = 2pf is the angular frequency (f is the frequency), s is arelaxation time, Cs is the steady state capacitance and C1 is thecapacitance taken at frequencies much larger than 1/(2pf) (Pethigand Kell, 1987; Kuang and Nelson, 1998). However, capacitance ef-fects are expected at the honey/electrode interfaces and this will af-fect the measured values. We have found that a modified version ofthe Cole–Cole equation (Cole and Cole, 1941, 1942) fits well thedata:

CðxÞ ¼ C1 þCs � C1

1þ ðxsÞa; ð2Þ

where a is an empirical parameter. The expression for the fit of con-ductance data was also chosen on the basis that it was the simplestmodel for conductance found in references Pethig and Kell (1987)and Kuang and Nelson (1998), based on the empirical Cole–Coleexpression:

GðxÞ ¼ G1 þ ðGs � G1ÞðxsÞa

1þ ðxsÞa: ð3Þ

In this expression the subscripts have the same meaning as before.Lastly, the expression used to model G/log(x) is purely empiri-

cal reflecting data shape and not a physical assumption:

GlogðxÞ ¼

xD

Aþ Bxþ Cffiffiffiffiffixp : ð4Þ

Example of fitting experimental data to expressions (2)–(4) areshown respectively in Fig. 1b (sample OB1, Al electrode), 1c (FL2/Pt) and 1d (AU4/Au). The fits were applied successively to eachsample, electrode and type of measure. Hence, each dataset of 21frequencies was substituted by simply four fit parameters. Thus,the initial 252 variables were replaced by 48 variables.

2.7.2. Fitting procedure for the Vis–NIR dataThe Vis–NIR spectra of the 13 samples are shown in Fig. 2a and

the fitting procedure to one of these spectra is exemplified inFig. 2c. Observation of Fig. 2a suggests that all curves have a linearslope between 750 and 850 nm. Hence, a baseline was determinedthrough a linear fit in this range and extrapolated across the entirewavelength. Three main peaks appear when the baseline is sub-tracted to the curves. These were fitted by Gaussian peaks.Fig. 2c shows a example of a fit. The Vis–NIR spectra were thenparameterized by eleven coefficients (two from the baseline and

three from each of the three Gaussians). A twelfth variable, thesquared sum of errors of the fit was then added.

2.7.3. Fitting procedure for the UV–Vis dataFig. 2b suggest a common baseline for all the spectra. It was

modelled as an inverse polynomial determined through a fit tothe first 10 points and to the last 50:

baseline ðkÞ ¼ 1Aþ Bkþ Ck2 þ Dk3 : ð5Þ

The spectra were then subtracted from the background and theresulting curve resembled a straight line with superimposed peaksand dips. The fit proceeded with two stages of gaussian peak/dip fit-ting. Each curve was then parameterized by a total of four parame-ters [the baseline] + (3 � 4) parameters [the four Gaussians] = 16parameters. Fig. 2d shows an example of this fit to sample SF2, withan almost perfect data fit.

2.7.4. Multi-way PCAEach one of the six datasets (raw data and fitting coefficients for

the e-tongue, Vis–NIR and UV–Vis spectroscopy measurements)constitutes a matrix to be analyzed through PCA. Scores corre-sponding to the first two principal components from each datasetwere stored in a new 6 � 2 matrix and subjected to new PCA.Group scaling (three groups) was used in the new matrix. Theresults of this final PCA were then analyzed as usual.

2.7.5. Ratio I/IThe ratio I/I (RII) was introduced in this work to help quantify-

ing the classification potential of each method. RII stands for the ra-tio of inter-distance to intra-distance (measured in the space ofprincipal components); that is, the ratio of the mean distance be-tween groups (inter-distance) to the mean distance between eachsample and the center of its group (intra-distance):

RII ¼ interdistanceintradistance

: ð6Þ

‘‘Groups’’ here mean the four honey groups. The center of eachgroup is simply the center of mass, denoted by an overbar, and isgiven by ðS1a; S2aÞ ¼ 1

Na

PNaia¼1ðS1aia ; S2aia Þ, where a = OB, AU, FL and

SF, S1 and S2 are the scores of the first and second principal compo-nents respectively, Na is the total number of samples in group a andia is the sample index within the group. Then the distance dab be-tween the centers of groups a and b isdab ¼ ½ðS1a � S1bÞ2 þ ðS2a � S2bÞ2�1=2 and the interdistance is definedas the average over all the dab values:

interdistance ¼ 16

X

a<b

dab; ð7Þ

where six corresponds to the number of combinations for fourgroups. The distance of each sample to the center of its group isdaia ¼ ½ðS1a � S1aia Þ

2 þ ðS2a � S2aia Þ2�1=2. The intradistance is defined

in this work as the average over the 13 daia values:

intradistance ¼ 113

X

a;ia

daia : ð8Þ

A high RII value indicates a separation between the groups that islarge compared to the sample dispersion within each group andhence good clustering. On the contrary, a small RII value indicatespoor clustering.

2.7.6. Variable selection through one-dimensional clusteringIt is possible to calculate the RII using only one variable. In this

case one retains only the first component in expressions (6)–(8).The meaning is simple: RII quantifies the clustering of the 13

a b

dc

Fig. 2. Optical spectroscopy data: (a) overlay plot of all the Vis–NIR spectra; (b) overlay plot of all the UV–Vis spectra; (c) example of a fit to a Vis–NIR spectrum throughbaseline and gaussian peak fitting; and (d) example of a fit to a UV–Vis spectrum through expressions (5) and gaussian peak fitting. The second stage peak fitting is notrepresented.

a b

c d

Fig. 1. Electronic tongue data: (a) overlay plot of all the data. Although not explicit in the graph (line coding degrades the graphics readability), the curves corresponding toAU honey tend to be on the top, and then, in descent order, SF (middle-top), OB (middle-bottom) and FL on the bottom (indicative trend); (b) example of a fit to a capacitancecurve through expression (2); (c) example of a fit to a conductance curve through expression (3); and (d) example of a fit to a G/log(x) curve through expression (4).

4 P.A. Ulloa et al. / Computers and Electronics in Agriculture 94 (2013) 1–11

samples in a one dimensional axis. We call variable selection by one-dimensional clustering the process of retaining the variables withhigher RII. Plots on the RII values for all the individual variablesused in this work are shown in Fig. 7.

2.7.7. PCA cross-validationA leave-one-out (LOO) cross-validation was performed on the

PCA models. One sample was excluded and a PCA model was ob-

tained for the remaining 12 samples. The sample left out was pro-jected on the PCA model to obtain its scores. Then, the distance ofthe sample left out to the center of each group was calculated fromits scores and assigned to the closest group. The process wasrepeated for all samples and the percentage of samples correctlyassigned through this process is represented by % assign.

P.A. Ulloa et al. / Computers and Electronics in Agriculture 94 (2013) 1–11 5

3. Results and discussion

3.1. Botanical origin of commercial honeys by a reference method

Standard melissopalynological results of the commercial hon-eys analyzed are summarized in Table 1. Honeys are typically clas-sified as unifloral when the frequency of the predominant plantflower pollen exceeds 45%. Under-representation (<45%) of thepredominant pollen was observed for some samples analyzed inthis study such as Citrus, Helianthus and Arbutus samples. Accord-ing to Von Der Ohe et al. (2004), the floral origin of a honey can beconsidered orange blossom (OB) if the sample contains 2–42% Cit-rus pollen content, and of strawberry-tree (AU) origin if the Arbu-tus pollen content ranges 8–20%. In the case of sunflower honey(SF), it should contain 12–92% Heliantus pollen grains while forFrench lavender honey (FL), samples must contain 10–57% Lavan-dula stoechas. Overall, data from the pollen analysis confirmedthe identity of the honey botanical source indicated by the pro-ducer on the product label. According to our results and the abovecriteria, the honeys analyzed in this study were classified with re-spect to their floral origin as follows: OB = Citrus honey, SF = sun-flower honey, FL = French lavender honey and AU = strawberry-tree honey.

3.2. Raw data of e-tongue, Vis–NIR and UV–Vis

The raw data obtained from the e-tongue is depicted in Fig. 1a.The 13 samples are shown simultaneously and the graph shows insequence the measurements of capacitance (C), conductance (G)and the ratio conductance over the log of frequency (G/log(x)),for each of the electrodes Al, Pt, Au and ITO. Although not explicitin the graph (line coding degrades the graphics readability), thehoney groups showed a clear offset relative to each other. Thecurves corresponding to AU honey tended to be on the top, andthen in descending order, SF, OB and FL on the bottom. This order-ing is only indicative and not followed by all individual samples.

The spectra obtained from Vis–NIR spectroscopy are shown inFig. 2a. All curves showed the highest absorbance level in the blueregion of the spectrum, suggesting the previously reported pre-dominant absorption in the UV range by the very abundant pheno-lic compounds (Antolovich et al., 2000; Tomás-Barberán et al.,2001). The AU samples exhibited the most distinctive spectra,showing higher plateaus along the visible region of the spectrum;SF samples showed the most pronounced decrease from 400 to500 nm; and OB and FL showed similar spectra. The Vis–NIR spec-tra show a consistent dip around 960 nm, probably due to a waterabsorption band near 970 nm (Hale and Querry, 1973).

The spectra obtained from UV–Vis spectroscopy are depicted inFig. 2b. All curves showed a tendency to increase towards the low-er limit of 200 nm, but the peak suggested by this tendency was

Table 1Pollen analysis of four randomly-selected commercial unifloral honey samples by the melishoney, FL = French lavender honey and AU = strawberry-tree honey (Arbutus unedo). Detailerapid screening was made to check conformity to the criteria described in Section 3.1.

Honey Pollinic spectrum (%)

OB Citrus spp. (30.2), Echium plantagineum (11.4), Rubus spp. (7.0), Cytisus scoparspp.(4.0), Lavandula stoechas (2.2), Crataegus monogyna (2.0), Raphanus spp./SAcacia spp. (1.0), Oxalis corniculata (0.6)

SF Helianthus annuus (24.6), Echium plantagineum (14.1), Pistacia terebinthus (6.9Rosmarinus officinalis (3.0), Eucalyptus spp. (3.0), Citrus spp. (2.9), Crepis capill(0.9), Lotus creticus (0.7), Reseda luteola (0.7), Crataegus monogyna (0.5)

FL Lavandula stoechas (54.2), Echium plantagineum (8.0), Rubus spp. (5.6), Cytisusspp. (1.4), Eucalyptus spp. (1.4), Rhamnus spp. (1.3), Quercus spp. (1.3), Senec

AU Arbutus unedo (32.0), Lavandula stoechas (15.1), Ceratonia siliqua (5.4), CytisuPistacia terebinthus (2.0), Olea europea (1.9), Calendula arvensis (1.5), Myrtus

out of the observable range. However, there were also well definedsecondary peaks. Peaks around 250 nm were associated with AUsamples while the peaks around 280 nm were mainly but not onlydue to SF samples. Otherwise, OB and FL were essentially indistin-guishable at first sight. The concentrations of the samples used forUV–Vis and Vis–NIR spectroscopy were very different and there-fore the absolute values of absorbance obtained in both methodscannot be compared directly.

3.3. Analysis of e-tongue data

Fig. 3 depicts the PCA and cluster analysis of the e-tongue data.The subplots on the left (a and c) refer to the full dataset prepro-cessed with first derivation (252 variables) and the subplots onthe right (b and d) refer to the analysis performed on the curve fit-ting coefficients (48 variables). Examination of Fig. 3a and b showsthat reasonable clustering was obtained in both cases, with slightlybetter results in Fig. 3b. Yet, in both cases sample AU2 was wronglyassociated with the SF cluster. The dendrograms obtained by clus-tering with K-Nearest-Neighbors (Fig. 3c and d) corroborated thePCA analysis. The bifurcation of the main branches in dendrogram3d took place at comparatively larger distances than in the case ofdendrogram 3c. Globally, this means that sample discriminationwas cleaner and more effective when using the curve fitting coef-ficients rather than the full dataset. Reinforcing this conclusion,RII = 5.5 in the latter while RII = 7.4 in the former case. However,under leave-one-out crossed validation, both methods gave thesame value % assign = 92% (12 in 13).

Despite the small number of samples (since the emphasis wasthe testing of the techniques), the results of this study comparewell with those from others. For example, the same ability to dis-tinguish between four types of Spanish honey, namely, three of flo-ral origin (Citrus, rosemary and polifloral) and one honeydew(forest origin) was reported in Escriche et al. (2012), who used apotentiometric e-tongue and PCA. Dias et al. (2008) used e-tongueand Linear Discriminant Analysis (LDA) to classify honey samplesof the most common three types (i.e., Erica spp., Echium spp. andLavandula spp. as predominant pollens) and obtained about 84%classification accuracy and 72% for cross-validation.

3.4. Analysis of Vis–NIR data

The results from the analysis of Vis–NIR data are depicted inFig. 4. The subplots on the left (a and c) refer to the full dataset pre-processed through a Savitzky–Golay filter of 95 points (each side),polynomial order 2 and derivative order 2 (3347 variables) and thesubplots on the right (b and d) refer to the analysis performed onthe curve fitting coefficients (12 variables). The results were gener-ally poor. In this case, and contrary to the e-tongue results, the PCAbased on the parameterization of the curves led to the worse re-

sopalynological standard method. OB = Citrus honey (Orange Blossom), SF = sunflowerd characterization was made only for these four samples. For the remaining samples a

ius (6.8), Pistacia terebinthus (5.1), Olea europea (4.8), Reseda luteola (4.4), Quercusinapsis spp.(1.7), Eucalyptus spp. (1.5), Rosmarinus officinalis (1.1), Erica spp. (1.1),

), Lavandula stoechas (5.4), Olea europea (5.3), Quercus spp. (5.1), Rubus spp. (3.5),aries (1.3), Campanula spp. (1.1), Senecio vulgaris (1.1), Raphanus spp./Sinapsis spp.

scoparius (5.5), Campanula spp. (2.9), Myrtus spp. (1.7), Olea europea (1.5), Acaciaio vulgaris (0.9)s scoparius (5.1), Echium plantagineum (4.5), Citrus spp. (3.9), Reseda luteola (3.9),spp. (1.4), Rosmarinus officinalis (0.9)

a b

dc

Fig. 3. Analysis of the e-tongue data. (a) PCA scores plot for the e-tongue data matrix processed with auto-scaling and first derivative; (b) PCA scores plot for the matrixobtained from the fitting coefficients to the e-tongue curves; (c) CA dendrogram corresponding to matrix used in a; and (d) CA dendrogram corresponding to matrix used in b.

a b

dc

Fig. 4. Analysis of the Vis–NIR data: (a) PCA scores plot for the Vis–NIR data matrix processed with auto-scaling and Savitzky–Golay filtering with second derivative; (b) PCAscores plot for the matrix obtained from the fitting coefficients to the Vis–NIR spectra; (c) CA dendrogram corresponding to matrix used in a; and (d) CA dendrogramcorresponding to matrix used in b.

6 P.A. Ulloa et al. / Computers and Electronics in Agriculture 94 (2013) 1–11

P.A. Ulloa et al. / Computers and Electronics in Agriculture 94 (2013) 1–11 7

sults. The cluster analysis (Fig. 4c and d) confirmed that with thefull dataset only one cluster (SF) was correctly isolated (Fig. 4c),while from the matrix of the fitting coefficients not a single clustercould be correctly identified (Fig. 4d). Consistently, RII = 3.2, for thefull dataset and RII = 1.8 for the fitting coefficients data. Under LOOcrossed validation the first method yielded a % assign = 62% (8 in13) and the second a % assign = 31% (4 in 13).

Two main comments are in order. First, the Vis–NIR spectrahave poor specificity, i.e., the patterns observed are caused bythe superposition of an indeterminate number of mainly unknownchemical components. Second, this poor specificity to the chemicalcomponents also implies a low specificity of the curve fittingprocedure.

Nevertheless, our results compare moderately well with previ-ous published results. In Corbella and Cozzolino (2005), Vis/NIRtogether with LDA and DPLS (discriminant partial least squares)models correctly classified, on average, more than 75% of the honeysamples belonging to pasture and more than 85% of the honeysamples belonging to Eucalyptus spp. Furthermore, NIR spectros-copy has also been applied successfully to discriminate honeysfrom several countries according to its botanical origin (Davieset al., 2002; Dvash et al., 2002; Corbella and Cozzolino, 2005;Woodcock et al., 2007, 2009). Depending on the models and themultivariate statistics used circa 80-90% accuracy in the discrimi-nation of honeys was attained.

3.5. Analysis of UV–Vis data

The results from the analysis of the UV–Vis data are depicted inFig. 5. The subplots on the left (a and c) refer to the full dataset pre-processed through a Savitzky–Golay filter of 27 points (each side),

a

dc

Fig. 5. Analysis of the UV–Vis data: (a) PCA scores plot for the UV–Vis data matrix procescores plot for the matrix obtained from the fitting coefficients to the UV–Vis spectracorresponding to matrix used in b.

polynomial order 2 and derivative order 2 (201 variables) and thesubplots on the right (b and d) refer to the analysis performed onthe curve fitting coefficients (16 variables). Once again the resultswere relatively poor and the comments of the last section concern-ing the Vis–NIR, apply here. The clustering was not clear except forthe AU group (Fig. 5b and d). Indeed, the fit coefficients matrixallowed a good separation of AU from the other groups, but apartfrom that no useful clustering was observed. The RII values werealso similar to those of the previous section: RII = 3.8 for the fulldataset and RII = 2.3 for the fitting coefficients data. Under leave-one-out crossed validation the first method yields a % assign = 69%(9 in 13) and the second a % assign = 46% (6 in 13).

3.6. Sensor fusion

The next step was to search for the best way to combine theresults of the different methods to improve the clustering ability.Fig. 6 shows the results obtained from data fusion performed intwo different ways. The subplots on the left (a and c) show theresults obtained from the simple concatenation of all the datasetsalready used (e-tongue, Vis–NIR and UV–Vis, both full and fit coef-ficients matrices) in a global matrix with a total of 3876 variables.One obvious conclusion from these results is that a simple piling ofdata does not mean better results. Indeed, the results from the glo-bal matrix were worse than the results obtained from the e-tonguedata alone. There was only one group correctly identified (SF,Fig. 6c) and all the indicators were poor: RII = 3.7 and %assign = 54% under LOO cross validation.

The subplots on the right (Fig. 6b and d) show the resultsobtained from MPCA. The idea behind the MPCA is to drop outall the unimportant information and keep the essential one in

b

ssed with auto-scaling and Savitzky–Golay filtering with second derivative; (b) PCA; (c) CA dendrogram corresponding to matrix used in a; and (d) CA dendrogram

PCA on all data MPCAa b

dc

Fig. 6. Data fusion – part I; (a) PCA scores plot for the matrix obtained by concatenating all the previous matrices; (b) final PCA scores plot obtained from multi-way PCA; (c)CA dendrogram corresponding to matrix used in a; and (d) CA dendrogram corresponding to matrix used in b.

8 P.A. Ulloa et al. / Computers and Electronics in Agriculture 94 (2013) 1–11

the form of the first two scores for each sample and for each one ofthe six datasets. The matrix used for MPCA had then 13 samplesand 12 variables, being expected that these 12 variables condensethe most important information of each method. Indeed, Fig. 6bshows a good clustering and solved the problem of misplacementof sample AU2 by the e-tongue analysis (Fig. 3b and d). The MPCA

Raw Data

e-to

ngue

Vis-

NIR

UV

a

c

e

Fig. 7. One dimensional clustering ability for each individual variable, as described by idata; (d) fitting coefficients d for Vis–NIR data; (e) UV data; (d) fitting coefficients d for

results were better than the results obtained from the individualmethods. The MPCA indicators were RII = 6.5 and % assign = 100%in LOO cross validation.

The MPCA approach seems, then, to be a good option for datafusion, although still rare in the honey research field. Hallis et al.(2010) reported on the classification of honey through fused data

Fit Coeffsb

d

f

ts RII value. (a) E-tongue data; (b) fitting coefficients for e-tongue data; (c) Vis–NIRUV data.

PCA on Partial Bests PCA on Global Bests

LOO cross validation PCADendrogram for

Partial Bests

a b

dc

Fig. 8. Data fusion – part II. Results obtained by variable selection through one-dimensional clustering (a) PCA scores plot for the matrix obtained by selecting the best 4–6variables in each of the six initial matrices (e-tongue, Vis–NIR and UV–Vis, both full datasets and fitting coefficient datasets) – method of Partial Bests; (b) PCA scores plot forthe matrix obtained by selecting the best 6 variables of all – method of Global Best; (c) leave-one-out cross validation for matrix used in a; (d) CA dendrogram correspondingto matrix used in a.

Table 2Analysis of 13 random commercial samples of four unifloral honeys by e-tongue, Vis–NIR and UV–Vis measurements.

Matrix nr. Type of data Nvara RIIb % assignc NC d Preprocess/methode

1 E-tongue full dataset 252 5.5 92 2 1st derivative2 Fitting coefficients of e-tongue data 48 7.4 92 2 curve fitting3 Vis–NIR full dataset 3347 3.2 62 1 Savitzky–Golay (191,2,2)4 Fitting coefficients of Vis–NIR data 12 1.8 31 0 Curve fitting5 UV–Vis full dataset 201 3.8 69 0 Savitzky–Golay (55,2,2)6 Fitting coefficients of UV–Vis data 16 2.3 46 1 Curve fitting7 Concatenation of all data 3876 3.7 54 1 All data above (matrices 1–6)8 MPCA 12 6.5 100 3 2 first scores from each matrix 1–69 Partial bests 28 10.6 100 4 Variable selection in each matrix 1–6

10 Global best 6 19.3 100 4 Variable selection from global matrix 7

a Nvar = number of variables.b RII = ratio interdistance to intradistance.c % assign = percent of correct cluster assignments in cross validation.d NC = Number of clusters correctly identified through cluster analysis.e Savistzky–Golay filters are characterized by three numbers: width (number of points), polynomial order and derivative order.

P.A. Ulloa et al. / Computers and Electronics in Agriculture 94 (2013) 1–11 9

of taste and smell. It was shown that fused data was better todetermine the floral origin of six varieties of honey from a localmanufacturer than when using taste or smell data by itself. Liet al. (2007) used the Cyranose 320™ electronic nose and thezNose™, which are two instruments used to detect volatile pro-files, to improve detection and classification performance of dam-aged apples. Similarly, a study that aimed to discriminate unifloralhoney from sugar syrup, and polifloral honey from sugar and adul-terated samples using the e-nose and e-tongue, exhibited a signif-

icant improvement in the capacity to separate the samples byapplying a sensor fusion technique (Zakaria et al., 2011).

3.7. Improvement of clustering through variable selection

MPCA keeps only the essential information contained in thefirst two PC’s of each matrix. However, the PC’s emerge from a cal-culation involving all the variables. In this sense, the effect of irrel-evant variables has not been fully avoided. It would be preferable

10 P.A. Ulloa et al. / Computers and Electronics in Agriculture 94 (2013) 1–11

to select the best variables a priori and run the PCA model onlyafter this variable selection. This variable selection was doneaccording to the procedure described previously in Section 2.7.6,which means that the variables with better one-dimensional clus-tering were chosen. The definition of RII was given in Section 2.7.5.

Fig. 7 shows the RII values for each individual variable. As statedin Section 2.7.5, a higher RII means a higher ability of a given var-iable to separate the honey groups all by itself. It may be observedthat the higher RII values appear within the e-tongue variables(plots a and b) and that a small range around 500 nm is the mostinteresting one in NIR (plot c). From the six graphs in Fig. 7 onepicks the variables with higher RII values, to build the eight andlast figure.

Fig. 8a shows the PCA scores plot obtained when one selectedthe best 4–6 variables (‘‘best’’ means higher RII, that is, betterone-dimensional clustering) in each matrix (e-tongue, Vis–NIRand UV–Vis, full and fit coefficients datasets) and concatenatesthe selected variables (for simplicity we will call this model ‘‘Par-tial Bests’’). The resulting matrix had 28 variables and showed avery good clustering ability. Indeed, all the groups were very wellseparated (Fig. 8a). This is consistent with a very high RII value(10.6) and a % assign = 100%. Fig. 8b shows the PCA scores plot ob-tained by selecting the best six variables from all the 3876 vari-ables available (this is the ‘‘Global Best’’ model). It turned outthat these best six variables came all from the e-tongue dataset.The clustering ability was also excellent (RII = 19.3), and %assign = 100%.

Fig. 8c shows the results of the LOO cross validation for themodel of Partial Bests. Each open symbol represents the left outsample in each validation round and the crosses represent the cor-responding calibration samples. It may be confirmed from this fig-ure that the left-out samples were all projected on the correctcluster. Finally, Fig. 8d shows the dendrogram obtained byclustering of the Partial Bests model. Table 2 summarizes the allthe results of this work.

In conclusion, the novel contributions from this work are thefollowing: (i) use of an impedance e-tongue based on a LCR meter,which is a simple, but new approach; (ii) demonstration of the fea-sibility of electronic e-tongue and optical spectroscopy data for thediscrimination of honey floral origin using MPCA; (iii) demonstra-tion of the advantage of using curve fitting parameters rather thanthe curves per se for PCA analysis; and (iv) introduction of a newmethod for variable selection, based on one-dimensional cluster-ing. This method improved significantly the classification results.

The emphasis of this work was the demonstration of thesemethods and we have used a restricted set of samples. Furtherwork with a larger number of samples must be undertaken to con-firm the results obtained by this investigation.

Acknowledgements

The authors would like to thank Professor Henrique Gomes,from the Instituto de Telecomunicações/Universidade do Algarvefor building the e-tongue and guide us through the measurements.We would like to thank also the careful revision of the manuscriptperformed by Professor J. Antonio Torres, from the Department ofFood Science & Technology of the Oregon State University. PabloUlloa was funded by Erasmus Mundus 2008-1022/002 EuropeanProgram Frame ECW/17. Ana M. Cavaco is an assistant researcherof CEOT, funded by FCT under Programa Ciência 2008-Contrataçãode Doutorados para o Sistema Científico e Tecnológico Nacional.Ana C. Figueira has been partially financed by Fundação para aCiência e a Tecnologia (FCT). CEOT, Programa Nacional de Re-equipamento Científico (PNRC) by Fundação para a Ciência e a

Tecnologia (FCT-REEQ/601/EEI/2005) and the POCI 2010-FEDERare also acknowledged.

References

Antolovich, M., Prenzler, P., Robards, K., Ryan, D., 2000. Sample preparation in thedetermination of phenolic compounds in fruits. Analyst 125, 989–1009.

Bianchi, F., Careri, M., Musci, M., 2005. Volatile norisoprenoids as markers ofbotanical origin of sardinian strawberry-tree (Arbutus unedo L.) honey:characterisation of aroma compounds by dynamic headspace extraction andgas chromatography-mass spectrometry. Food Chemistry 89, 527–532.

Cavaco, A.M., Miguel, G., Antunes, D., Guerra, R., 2012. Chapter 9: Determination ofgeographical and botanical origin of honey: from sensory evaluation to the stateof the art of non-invasive technology. In: Bondurand, G., Bosch, H. (Eds.), Honey:Production, Consumption and Health Benefits. Nova Science Publishers, NY,USA, pp. 1–40.

Codex Alimentarius Commission, 2001. Revised Codex Standard for Honey. CodexAlimentarius Commission. Codex Stan, 12-1981, Rev. 1 (1987), Rev. second ed.

Cole, K.S., Cole, R.H., 1941. Dispersion and absorption in dielectrics I. Alternatingcurrent characteristics. Journal of Chemical Physics 9, 341–351.

Cole, K.S., Cole, R.H., 1942. Dispersion and absorption in dielectrics II. Direct currentcharacteristics. Journal of Chemical Physics 10, 98–105.

Corbella, E., Cozzolino, D., 2005. The use of visible and near infrared spectroscopy toclassify the floral origin of honey samples produced in Uruguay. Journal of NearInfrared Spectroscopy 13, 63.

Davies, A., Radovic, B., Fearn, T., Anklam, E., 2002. A preliminary study on thecharacterisation of honey by near infrared spectroscopy. Journal of NearInfrared Spectroscopy 10, 121.

Dias, L.A., Peres, A.M., Vilas-Boas, M., Rocha, M.A., Estevinho, L., Machado, A.A.S.C.,2008. An electronic tongue for honey classification. Microchimica Acta 163, 97–102.

Dvash, L., Afik, O., Shafir, S., Schaffer, A., Yeselson, Y., Dag, A., Landau, S., 2002.Determination by near-infrared spectroscopy of perseitol used as a markerfor the botanical origin of avocado (Persea americana mill.) honey. Journal ofAgricultural and Food Chemistry 50, 5283–5287.

Escriche, I., Kadar, M., Domenech, E., Gil-Sánchez, L., 2012. A potentiometricelectronic tongue for the discrimination of honey according to the botanicalorigin. Comparison with traditional methodologies: physicochemicalparameters and volatile profile. Journal of Food Engineering 109, 449–456.

Estevinho, L., Pereira, A.P., Moreira, L., Dias, L.G., Pereira, E., 2008. Antioxidant andantimicrobial effects of phenolic compounds extracts of northeast Portugalhoney. Food and Chemical Toxicology 46, 3774–3779.

European Union, 2001. Council Directive 2001/110 EC Relating to Honey. EuropeanUnion, Official Journal of the European Communities.

Hale, G.M., Querry, M.R., 1973. Optical constants of water in the 200-nm to 200-lmwavelength region. Applied Optics 12, 555–563.

Hallis, A., Ammar, A., Yeon, A., Shakaff, M., Hamid, A., Jamilah, M., Masnaet, A., 2010.Determination of honey floral origin using data fusion of electronic taste andsmell sensing system. In: Proceedings of the 1st Regional Conference on Appliedand Engineering Mathematics (RCAEM-I), pp. 183–184.

Hruskar, M., Krpan, M., Markovic, K., Saric, G., Major, N., Vahcic, N., Curic, D., 2008.Applications of electronic tongues in determination of honey origins. In:Proceedings of the 2008 Joint Central European Congress, 4th Central EuropeanCongress on Food, 6th Croatian Congress of Food Technologists Biotechnologistsand Nutritionists, Cavtat, Hrvatska.

Kaškoniene, V., Venskutonis, P.R., 2010. Floral markers in honey of various botanicaland geographic origins: a review. Comprehensive Reviews in Food Science andFood Safety 9, 620–634.

Kuang, W., Nelson, S., 1998. Low-frequency dielectric properties of biologicaltissues: a review with some new insights. Transactions of the ASAE 41, 173–184.

Li, C., Heinemann, J., Irudayaraj, P.H., 2007. Detection of apple deterioration using anelectronic nose and znose™. Transactions of the ASABE 50, 1417–1425.

Lieux, M.H., 1980. Acetolysis applied to microscopical honey analysis. Grana 19, 57–61.

Louveaux, J., Maurizio, A., Vorwohl, G., 1978. Methods of melissopalynology. BeeWorld 59, 139–157.

Nicolai, B.M., Beullens, K., Bobelyn, E., Peirs, A., Saeys, W., Theron, K.I., Lammertyn, J.,2007. Nondestructive measurement of fruit and vegetable quality by means ofNIR spectroscopy: a review. Postharvest Biology and Technology 46, 99–118.

Ozer, N., Engel, B., Simon, J., 1995. Fusion classification techniques for fruit quality.Transactions of the ASABE 38, 1927–1934.

Pethig, R., Kell, D.B., 1987. The passive electrical properties of biological systems:their significance in physiology, biophysics and biotechnology. Physics inMedicine and Biology 32, 933.

Tomás-Barberán, F.A., Martos, I., Ferreres, F., Radovic, B.S., Anklam, E., 2001. HPLCflavonoid profiles as markers for the botanical origin of European unifloralhoneys. Journal of the Science of Food and Agriculture 81, 485–496.

Valdés, B., Diez, M., Fernandez, I., 1987. Atlas polnico de Andalucá Occidental.Technical Report. nstituto de Desarrollo Regional No. 43. Universidad de Sevilla.EXCMA Diputation de Cadiz, Spain.

Von Der Ohe, W., Oddo, L.P., Piana, M.L., Morlot, M., Martin, P., 2004. Harmonizedmethods of melissopalynology. Apidologie, 35.

Wei, Z., Wang, J., Liao, W., 2009. Technique potential for classification of honey byelectronic tongue. Journal of Food Engineering 94, 260–266.

P.A. Ulloa et al. / Computers and Electronics in Agriculture 94 (2013) 1–11 11

Woodcock, T., Downey, G., Kelly, J.D., O’Donnell, C., 2007. Geographical classificationof honey samples by near-infrared spectroscopy: a feasibility study. Journal ofAgricultural and Food Chemistry 55, 9128–9134.

Woodcock, T., Downey, G., O’Donnell, C.P., 2009. Near infrared spectralfingerprinting for confirmation of claimed PDO provenance of honey. FoodChemistry 114, 742–746.

Zakaria, A., Shakaff, A.Y.M., Masnan, M.J., Ahmad, M.N., Adom, A.H., Jaafar, M.N.,Ghani, S.A., Abdullah, A.H., Aziz, A.H.A., Kamarudin, L.M., Subari, N., Fikri, N.A.,2011. A biomimetic sensor for the classification of honeys of different floralorigin and the detection of adulteration. Sensors 11, 7799–7822.