Chemometric analysis of French lavender and lavandin essential oils by near infrared spectroscopy

HAL Id: hal-01451404

https://hal.archives-ouvertes.fr/hal-01451404

Submitted on 12 Apr 2018

HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.

L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.

Chemometric analysis of French lavender and lavandin essential oils by near infrared spectroscopy

Sofia Lafhal, Pierre Vanloot, Isabelle Bombarda, Jacky Kister, Nathalie Dupuy

To cite this version:

Sofia Lafhal, Pierre Vanloot, Isabelle Bombarda, Jacky Kister, Nathalie Dupuy. Chemometric analysis

of French lavender and lavandin essential oils by near infrared spectroscopy. Industrial Crops and

Products, Elsevier, 2016, 80, pp.156 - 164. �10.1016/j.indcrop.2015.11.017�. �hal-01451404�

ContentslistsavailableatScienceDirect

Industrial Crops and Products

j ourna l h o m e pa g e : w w w . e l s e v i e r . c o m / l o c a t e / i n d c r o p

Chemometric analysis of French lavender and lavandin essential oils by near infrared spectroscopy

Sofia Lafhal

, Pierre Vanloot, Isabelle Bombarda, Jacky Kister, Nathalie Dupuy

AixMarseilleUniversité,LISA,EA4672,EquipeMETICA,13397Marseillecedex20,France

a r t i c l e i n f o

Articlehistory:

Received22July2015

Receivedinrevisedform4November2015 Accepted6November2015

Availableonline6December2015

Keywords:

Essentialoils Lavender Lavandin Metabolomics NIR Chemometrics

a b s t r a c t

Chemometrictreatmentsofnearinfrared(NIR)spectrawereusedfirstlytounderstanddatastructureby principalcomponentanalysis(PCA),todiscriminate,bypartialleastsquares-discriminantanalysis(PLS- DA)regression,Frenchlavenderandlavandinessentialoils(EOs)samples(n=160)andthesevenvarieties (Abrial,Fine,Grosso,Maillette,Matherone,SumianandSuper)andtoquantifythemaincompoundssuch aslinalylacetate,linalool,eucalyptolandcamphorbyPLSregressionmodels.Thestudywascarried outoverthreecropyears(2012–2014)totakeseasonalvariationsintoaccount.Frenchlavenderand lavandinEOsandtheirvarietieswerewellclassified(100%forlavender/lavandinEOsandbetween96 and100%forvarieties)byPLS-DAregressionmodels.ThecalibrationmodelsobtainedbyPLSregression forthedeterminationofthemaincompoundcontentsrevealedgoodcorrelation(≥0.97)betweenthe predictedandreferencevalues.Inthecaseofmajorcompoundsincludinglinalylacetateandlinalool, therelativeerrorofprediction(REP)iscloseto2.5%.Partialleastsquaresregressionvectorsallowed ustoidentifylavandulylacetate,eucalyptol,linalool,camphor,trans-␤-ocimene,␤-caryophylleneand linalylacetateasmetabolomicindicatorsofFine,Maillette,Matherone,Abrial,Grosso,SuperandSumian varietiesrespectively.TheuseofNIRspectraallowedforanimprovementinFrenchlavenderandlavandin EOscharacterization,qualitycontrolandtraceability.

©2015ElsevierB.V.Allrightsreserved.

1. Introduction

Thelavandulafamily (Lamiaceae)is composedofthirty nine lavandulaspecieswhicharemostlyofMediterraneanorigin,suchas Lavandulaangustifolia(lavender)andnumerousintraspecifictaxa andhybridssuchaslavandin. Thesmellofthelavandula family anditscoloraretypicaloftheMediterraneanlandscape,partic- ularlyinFrance.Itsessentialoilisusedforcosmetics,perfumes, andinmedicine(Cawthorn,1995;KnowltonandPearce,1993;Lis- Balchin,2003;Lis-Balchinand Hart,1999;Piccagliaetal.,1993;

Raut and Karuppayil, 2014; Vakili et al., 2014).The Lamiaceae familyis classifiedin severalspecies whichare subdividedinto varieties,eachwithitsphysicalandenvironmentalcharacteristics.

ThebestknownisthepopulationlavenderincludingL.angustifo- lia,andamongitsvarieties(Maillette,MatheroneandFine),the Finevarietyisthemostfamousthankstoitsessentialoilyieldand itshighlinaloolcontent.LavandinisahybridbetweenL.angusti- foliaP.Mill.andL.latifolia(L.f.)Medikus,whosecultureismore recentandmorefocusedon“industrial”production.Amongthe

∗Correspondingauthor.

E-mailaddress:sof.lafhal@gmail.com(S.Lafhal).

mainvarietiesoflavandinEOs(Abrial,Grosso,SumianandSuper), theGrossovarietyisthemostfamousforitsessentialoilyield(Lis- Balchin,2004).Lavenderandlavandinoilscontainmorethanone hundredcompoundsincludinglinalylacetate,linalool,camphor, borneol,eucalyptoland␤-caryophyllene,eachcontributingtothe chemical and sensoryproperties of theoils. Themajor distinc- tionbetweenvarietiesoflavandinessentialoilsisintheirrelative contentsoflinalylacetate,linalool,eucalyptolandcamphor.The chemicalcompositionoflavandinhasahighercamphorcontent thantruelavender.Chemicalcompositioncanberevealedusinggas chromatography(Bicchi,2000;CanBas¸erandÖzek,2012;Cserháti etal., 2005;Daferera etal., 2002;Marriott etal.,2001; Nikoli ´c etal.,2014;Palladoetal.,1997;TerHeideetal.,1970)andgas chromatography-sniffing(ChinandMarriott,2015;Cserhátietal., 2005;Marriottetal.,2001).Thesemethodsareusuallyappliedfor qualitycontrolpurposes(Doetal.,2015)andselection ofhigh- qualityplants,but theyarevery time-consuming,and attempts havebeenmadetofindalternativeanalyticalanalysismethods.

Also, toourknowledge, thequality ofan essential oildepends onthreemainfactors:qualityoftheplant,harvesting,distillation.

However,onthebasisofthesecriteria,onlyspecialistsareableto differentiateonevarietyfromanother.Inthiscontext,vibrational spectroscopicmethodssuchasNIRspectroscopyincombination http://dx.doi.org/10.1016/j.indcrop.2015.11.017

0926-6690/©2015ElsevierB.V.Allrightsreserved.

withchemometrictreatmentsweresuccessfullyintroducedfora non-destructivedeterminationofmetabolitesoccurringinessen- tialoils,andtofacilitatetheimplementationofthismethodonan industrialscale.Vibrationalspectroscopymethodsofferaglobal metabolicfingerprint of essential oils and allow theprediction of metabolite contents suchas camphor (Allwoodet al., 2007;

Bombardaetal.,2008;Cozzolino,2009;Dupuyetal.,2013,2014;

Maiettietal.,2013;Mocoetal.,2007;Schulzetal.,2008;Tankeu etal.,2014).

TheaimofthisstudywastoshowtheadvantagesofNIRspec- troscopyassociatedwithchemometrictreatmentsfordirectand rapidtestmethods.ThesecanbeusedforlavenderandlavandinEOs characterization,qualitycontrolandtraceability.Thismethodology canalsobeusedtoidentifymetabolomicindicatorsandestablish afasterqualitycontrolfordiscriminationoflavenderandlavandin essentialoilsbasedontheirspecificities.

Toachievethesegoals,inthefirstpartofthisstudy,datamining wascarriedoutbyprincipalcomponentanalysis(PCA)toevaluate differencesandsimilaritiesbetweensamplespectra;partialleast squares(PLS)regressionwasusedtodeveloplavender/lavandin EOs orvarietiesprediction models.Partialleastsquares regres- sionwasalsousedtoquantifythemaincompoundssuchaslinalyl acetate,linaloolandcamphorwithgaschromatographyasrefer- encedata.Inthesecondpart,thePLSregressionvectorsobtained forthepredictionofthemain compoundswerecompared with thePLSregressionvectorsobtainedforthepredictionoftheEO varietiestoidentifythemetabolomicindicatorsofeachvariety.

2. Materialsandmethods

2.1. Essentialoilsamples

Onehundredsixtysampleswereanalyzedincludinglavandin oil samples (n=94) and lavender oil samples (n=66) obtained from three French cooperatives (“Société Coopérative Parfums ProvenceVentoux”inSault,“SociétéCoopérativedesPlantesàPar- fumsdeProvence”inSimianelaRotondeand“France Lavande”

inMontguers),over threeyearsofharvest(2012–2014),includ- ingvariousvarietiesandvariousFrenchcollectareas(Unknown department (00), Alpes-de-Haute-Provence (04), Ardèche (07), Drôme(26)andVaucluse (84)).Thelavender/lavandinessential oilsstudiedwere[L.angustifoliaMiller](lavender)anditshybrid [L.angustifoliaMiller×L.latifoliaLinnaeusfilsMedikus](lavandin) typeFrance;theyweredividedintovarieties:Fine(FI,n=21),Mail- lette(MA,n=29)andMatherone(MT,n=16)forlavendersamples andAbrial(AB,n=16),Grosso(GR,n=37),Sumian(SU,n=19)and Super(SP,n=22)forlavandinsamples.

2.2. Purestandardsamples

Pure standard substances linalyl acetate, linalool, camphor, eucalyptol,borneol,trans-␤-ocimene, lavandulylacetate and␤- caryophyllenewerepurchasedfromLavenderFrance(Montguers, France), Fluka (Buchs, Switzerland), Alpha Aesar (Karlsruhe, Germany),Merck(Schuchardt,Germany),AlphaAesar(Karlsruhe, Germany),Sigma–Aldrich(Steinheim,Germany),Adrian(Aix-les- Milles,France)andTCIEurope(Zwijndrecht,Belgium)respectively.

2.3. Gaschromatography(GC)

2.3.1. Gaschromatographycoupledtoamassspectrometer (GC–MS)analysis

Gaschromatography coupledtoa massspectrometeranaly- seswereperformedona7890AGCsystemcoupledtoa 5975C VLmassspectrometerdetector(AgilentTechnologies)equipped withaHP-5MScapillarycolumn(J&WScientific,30m×0.25mm,

0.25␮mfilmthickness).Dataacquisitionandprocessingwereper- formedusingtheMSDChemstationE.01.01.335(Agilent)software.

Onemicroliterofdilutedessentialoil(80␮Lin1.5mLofethanol) wasinjected.Theexperimentalconditionsdevelopedinthelabo- ratorywere:oventemperatureprogram,2minat80^◦C,then80^◦C to200^◦C(5^◦C/min),then200^◦Cto260^◦C(20^◦C/min),andheldat finaltemperaturefor5min;temperaturesinjector(splitratio60) anddetectorweresetto250^◦C;carriergaswasheliumataflow rateof1.2mL/min;solventdelay2min;ionizationvoltage70eV;

electronmultiplier1kV.

2.3.2. Gaschromatography(GC)analysis

Gas chromatography analyses were performed on a 7890A GC(AgilentTechnologies)systemwithaflameionisationdetec- tor(FID)equippedwithaHP5 capillarycolumn(J&WScientific, 30m×0.25mm,0.25␮mfilmthickness).Thedataacquisitionand processingwereperformedusingtheChemstationB.04.03-SP1(87) (Agilent)software.Theexperimentalconditionsdevelopedinthe laboratorywerethesameasGC–MSexceptforcarriergaswhich washydrogen.Linearretentionindiceswerecalculatedwithrefer- enceton-alkanes(C8–C28).

2.4. Nearinfrared(NIR)spectroscopy

ThespectraofeachlavenderorlavandinEOsandliquidstan- dard compounds placed in a quartzcell (2mm)wererecorded from4000to10,000cm⁻¹usingthesoftwareresultintegration2.1 (ThermoNicolet),with4cm⁻¹resolutionand64scansonaNico- letAntarisIIspectrometerequippedwithanInGaAsphotodiode detector,anH₂ NIRsourceandaCaF₂-germaniumbeamsplitter.

Anemptyquartzcellwastakenasreferenceforthebackground spectrumbeforeanalysisofeachsample.Thenearinfraredspec- trometerwassituatedinanair-conditionedroom(21^◦C).Spectra ofsolidstandardcompounds(camphorandborneol)wererecorded usinganintegratingsphereindiffusereflectancemodefrom4000 to10,000cm⁻¹,with4cm⁻¹resolutionand64scans.Abackground spectrumwascollectedunderthesameconditionsbeforemeasure- mentofeachsample(interleavedmode).

2.5. Chemometricanalysis

2.5.1. Principalcomponentanalysis(PCA)

Principal component analysis (PCA) (Cozzolino et al., 2011;

Esbensenetal.,2002;Kumaretal.,2014;Woldetal.,1987)isan unsupervisedmodelingmethodthatallowsforexploratorydata analysis;itextractsinformationfromdatasetandremovesnoise;it reducesthenumberofdimensions;anditallowsforclassificationof samplesbyinvestigatingsimilaritiesanddifferencesbetweenthe samples.Theprincipalcomponentanalysisprojectsintoasmaller numberoflatentvariablescalledprincipalcomponents(PC).Each principalcomponentexplainspartofthetotalinformation con- tainedintheoriginaldataandthefirstPCistheonethatcontains themostinformation,followedindescendingorder intermsof informationbyPC2andPC3andsoon.PlottingtwoPCsrelatively toeachotherthusallowsforinterpretationofsomegroups,thanks tothesimilaritiesordifferencesbetweensamples.

2.5.2. Partialleastsquaresregression(PLS)

Partialleastsquares(PLS)regression(Daszykowskietal.,2007;

Esbensen et al.,2002; Kumaret al.,2014; Liangand Kvalheim, 1996;Martens,1979;Sjöströmetal.,1983)isasupervisedmethod whichisbasedontherelationbetweensignalintensity(spectrum) andthecharacteristicsofthesample(Yvariable).Interferenceand overlapping informationmaybeovercomebyusing apowerful multicomponentanalysissuchasPLS.Thealgorithmisbasedon theabilitytomathematicallycorrelatespectraldatatoaproperty

Table1

ChemicalcompositionoflavenderEOvarieties(GC%).

RI^a Compounds FI(n=21) MA(n=29) MT(n=16) Lavender(n=66)

Mean SD Mean SD Mean SD Mean SD

923 ␣-Pinene 0.16 ±0.06 0.07 ±0.04 0.08 ±0.05 0.10 ±0.06

944 Camphene 0.10 ±0.05 0.20 ±0.07 0.07 ±0.04 0.12 ±0.08

973 Sabinene 0.35 ±0.12 0.41 ±0.14 0.32 ±0.12 0.36 ±0.14

980 ␤-Pinene 0.07 ±0.08 0.02 ±0.04 0.02 ±0.03 0.04 ±0.06

985 3-Octanone 0.78 ±0.38 1.37 ±0.36 0.36 ±0.15 0.84 ±0.52

989 ␤-Myrcene 0.39 ±0.14 0.34 ±0.19 0.42 ±0.12 0.38 ±0.16

1012 Hexylacetate 0.26 ±0.07 0.36 ±0.13 0.09 ±0.11 0.24 ±0.15

1030 Limonene 0.32 ±0.13 0.15 ±0.06 0.11 ±0.11 0.19 ±0.14

1034 Eucalyptolandcis-␤-Ocimene 4.13 ±0.79 1.43 ±0.63 5.95 ±1.35 3.84 ±2.07

1044 trans-␤-Ocimene 2.95 ±0.60 1.02 ±0.64 7.42 ±1.67 3.80 ±2.72

1073 Linalooloxide 0.16 ±0.04 0.38 ±0.08 0.12 ±0.03 0.22 ±0.14

1087 ␣-Terpinolene 0.12 ±0.10 0.22 ±0.17 0.07 ±0.08 0.14 ±0.15

1098 Linalool 26.92 ±3.47 39.09 ±6.55 19.74 ±3.55 28.58 ±9.50

1106 Octen-1-olacetate 0.97 ±0.28 0.72 ±0.30 1.02 ±0.29 0.90 ±0.32

1145 Hexylisobutyrate 0.08 ±0.04 0.06 ±0.03 0.02 ±0.03 0.05 ±0.04

1150 Camphor 0.33 ±0.17 0.56 ±0.17 0.27 ±0.08 0.39 ±0.20

1168 Lavandulol 0.98 ±0.38 0.18 ±0.29 1.53 ±1.11 0.90 ±0.83

1171 Borneol 1.11 ±0.21 1.72 ±0.34 0.83 ±0.22 1.22 ±0.47

1183 Terpinen-4-ol 4.12 ±1.54 0.38 ±0.33 1.86 ±0.40 2.12 ±1.85

1193 Hexylbutyrate 0.34 ±0.07 0.53 ±0.11 0.11 ±0.13 0.33 ±0.20

1197 ␣-Terpineol 0.74 ±0.32 0.72 ±0.33 0.77 ±0.43 0.74 ±0.35

1260 Linalylacetate 38.22 ±2.72 40.01 ±5.32 39.98 ±2.26 39.40 ±4.09

1293 Lavandulylacetate 4.60 ±1.66 0.71 ±0.43 5.53 ±1.36 3.61 ±2.46

1386 Geranylacetate 0.55 ±0.18 0.54 ±0.21 0.48 ±0.22 0.52 ±0.21

1429 ␤-Caryophyllene 5.09 ±0.71 3.67 ±0.89 6.12 ±0.71 4.96 ±1.28

1460 ␤-Farnesene 1.76 ±0.47 1.85 ±0.26 2.92 ±0.48 2.18 ±0.62

1492 GermacrenD 0.83 ±0.24 0.27 ±0.11 0.69 ±0.13 0.60 ±0.30

1594 Caryophylleneoxide 0.53 ±0.15 0.48 ±0.14 0.50 ±0.18 0.50 ±0.16

aRI:RetentionindicesonHP-5capillarycolumnandSD:Standarddeviation.

Fig.1. NIRspectraoflavender(FI)andlavandin(AB)essentialoils.

matrixofinterestwhilesimultaneouslyaccountingforallother significantspectralfactorsthatperturbthespectrum.Itisthusa multivariateregressionmethodthatusesthefullspectralregion selectedandisbasedontheuseoffactors.Theevaluationoferrors incalibrationwascarriedoutbycomputingthestandarderrorof calibration(SEC)asfollows:

SEC=

⎛

⎜ ⎝

N

i=1(Ci−Ci)² N−1−p

⎞

⎟ ⎠

whereC_iistheknownvalue,C’_iisthevaluecalculatedbythecali- brationequation,Nisthenumberofsamplesandpisthenumber

of independentvariables in theregression optimized by cross- validation.

Thestandarderrorofprediction(SEP)givesanestimationof thepredictionperformance duringthestepofvalidation ofthe calibrationandiscalculatedfromthefollowingequation:

SEP=

⎛

⎜ ⎝

M

i=1(Ci−Ci)² M

⎞

⎟ ⎠

whereMisthenumberofsamplesinthepredictionset.

Table2

ChemicalcompositionoflavandinEOvarieties(GC%).

RI^a Compounds AB(n=16) GR(n=37) SP(n=19) SU(n=22) Lavandin (n=94)

Mean SD Mean SD Mean SD Mean SD Mean SD

923 ␣-Pinene 0.30 ±0.09 0.32 ±0.10 0.11 ±0.03 0.22 ±0.09 0.24 ±0.12

944 Camphene 0.27 ±0.06 0.22 ±0.06 0.14 ±0.04 0.20 ±0.06 0.21 ±0.07

973 Sabinene 0.59 ±0.15 0.36 ±0.10 0.13 ±0.05 0.27 ±0.08 0.34 ±0.18

980 ␤-Pinene 0.33 ±0.14 0.30 ±0.09 0.04 ±0.04 0.15 ±0.17 0.20 ±0.16

985 3-Octanone 0.22 ±0.07 0 ±0 0.68 ±0.18 0.86 ±0.36 0.44 ±0.41

989 ␤-Myrcene 0.42 ±0.06 0.48 ±0.10 0.51 ±0.14 0.43 ±0.09 0.46 ±0.11

1012 Hexylacetate 0.16 ±0.05 0.13 ±0.04 0.41 ±0.17 0.04 ±0.05 0.18 ±0.16

1030 Limonene 0.67 ±0.09 0.62 ±0.10 0.70 ±0.40 0.87 ±0.21 0.71 ±0.24

1034 Eucalyptolandcis-␤-Ocimene 8.41 ±0.98 5.23 ±0.97 3.55 ±0.46 7.52 ±1.96 6.18 ±2.10

1044 trans-␤-Ocimene 2.89 ±0.55 0.29 ±0.06 1.49 ±0.27 1.72 ±0.63 1.60 ±1.01

1073 Linalooloxide 0.14 ±0.02 0.12 ±0.03 0.11 ±0.03 0.13 ±0.02 0.12 ±0.03

1087 ␣-Terpinolene 0.36 ±0.03 0.35 ±0.04 0.26 ±0.03 0.34 ±0.04 0.33 ±0.05

1098 Linalool 34.72 ±2.00 33.79 ±2.55 35.32 ±2.63 43.13 ±3.56 36.74 ±4.47

1106 Octen-1-olacetate 0.48 ±0.11 0.31 ±0.07 0.28 ±0.07 0.19 ±0.05 0.31 ±0.12

1145 Hexylisobutyrate 0.18 ±0.01 0.19 ±0.02 0.14 ±0.02 0.19 ±0.03 0.17 ±0.03

1150 Camphor 9.18 ±0.56 6.90 ±0.63 4.70 ±0.44 6.07 ±1.28 6.71 ±1.63

1168 Lavandulol 0.72 ±0.20 0.68 ±0.26 0.38 ±0.22 0.22 ±0.32 0.50 ±0.32

1171 Borneol 2.96 ±0.44 3.11 ±0.56 2.92 ±0.49 6.20 ±1.69 3.80 ±1.56

1183 Terpinen-4-ol 0.91 ±0.29 3.50 ±0.68 0.30 ±0.21 0.72 ±0.75 1.36 ±1.53

1193 Hexylbutyrate 0.39 ±0.06 0.40 ±0.06 0.69 ±0.12 0.52 ±0.11 0.50 ±0.15

1197 ␣-Terpineol 0.78 ±0.15 0.93 ±0.32 0.99 ±0.36 0.96 ±0.17 0.91 ±0.29

1260 Linalylacetate 25.47 ±1.96 31.13 ±2.60 38.02 ±2.68 21.72 ±3.24 29.08 ±6.33

1293 Lavandulylacetate 1.55 ±0.14 2.38 ±0.34 1.49 ±0.30 0.60 ±0.44 1.50 ±0.74

1386 Geranylacetate 0.44 ±0.07 0.51 ±0.14 0.63 ±0.16 0.46 ±0.08 0.51 ±0.14

1429 ␤-Caryophyllene 2.62 ±0.27 1.76 ±0.17 1.36 ±0.14 1.43 ±0.38 1.79 ±0.49 1460 ␤-Farnesene 0.83 ±0.16 1.37 ±0.18 0.84 ±0.09 1.00 ±0.18 1.01 ±0.29

1492 GermacrenD 0.71 ±0.09 0.75 ±0.11 0.57 ±0.08 0.53 ±0.12 0.64 ±0.14

1594 Caryophylleneoxide 0.12 ±0.03 0.08 ±0.02 0.09 ±0.03 0.08 ±0.05 0.09 ±0.04

aRI:RetentionindicesonHP-5capillarycolumnandSD:Standarddeviation.

Anotheruseful parameter is the relative error of prediction (REP),whichshowsthepredictiveabilityofthemodel,calculated fromtheequation:

REP=SEP

y ×100 (3)

where⁻yisthemeanoftheobservedvaluesofcompoundcontent.

Partialleastsquaresregressionhasnotbeendevelopedforpat- tern recognitionproblems such as classification. However, this techniquecanbeadaptedforclassification,givingrisetothepar- tialleastsquare-discriminantanalysis(PLS-DA)regressionmethod.

Partialleastsquare-discriminantanalysisregressioniscarriedout usinganexclusivebinarycoding schemewithonebitperclass (lavender/lavandinorvariety).Forthecodificationofsamples,the twoEOscorrespondingtolavenderandlavandinwerearbitrarily classifiedinthatorder.Forinstance,alavendersamplewascodi- fiedbythevector{1;0}.Thesamplewasthenassignedtotheclass showingthehighestmembershipvalue.Becauseofthedifficultyof calibratingandpredictingsampleswithbinaryvariablesitisneces- sarytograderesultsbetweenvalues0and1.Samplevalueslower than0.5andhigherthan1.5wereidentifiedasoutsidethedefined classandsampleswithvaluesbetween0.5and1.5wereidentified asbelongingtothedefinedclass.Thesameprotocolwasusedto predictvarieties.Forthecodificationofsamples,thesevenvarieties correspondingtoAB,GR,SU,SP,FI,MAandMTwerearbitrarilyclas- sifiedinthatorder.Forinstance,anABsamplewascodifiedbythe vector{1;0;0;0;0;0;0}.

AsHaalandandThomas(1988)explained,thefirstcoefficientof regression(notedB)isagoodapproximationofthepurecompound spectruminthecaseofPLSregression.Ifweextendthisproperty ofthefirstregressioncoefficienttotheclasspredictionbyPLS-DA, thefirstcoefficientofregressionisanapproximationoftheorig- inalfeatureofthecomplexchemicalsystemrepresentingaclass.

Inthecontextofourstudy,thoseregressioncoefficientsBwere veryusefultoidentifymetabolomicindicatorswhicharetheorig-

inalfeaturesofthelavenderandlavandinEOs.Thefirstregression coefficientsBobtainedforeachEOvarietiespredictionwerecom- paredwiththoseobtainedforthemaincompoundsprediction.The metabolomicindicatorofavarietywasthemajorcompoundwith thefirstregressioncoefficientthemostsimilartothatobtained forthepredictionofthevariety.Theevaluationofthesimilarity betweenregressioncoefficientswascarriedoutbycomputingthe Pearsoncoefficientasfollows:

r= COV(X,Y)

varX×varY (4)

wherecov(X,Y)isthecovarianceandvaristhevariance.

TobuildallPLSandPLS-DAmodels,107sampleswereused:

(FI,n=14),(MA,n=19),(MT,n=11),(AB,n=11),(GR,n=25),(SU, n=13)and(SP,n=14)andthevalidationstepwasperformedby fullcross-validation.Totesttheperformanceofthemodelsinpre- diction,53sampleswereused:(FI,n=7),(MA,n=10),(MT,n=5), (AB,n=5),(GR,n=12),(SU,n=6)and(SP,n=8).Forbestprediction performance,inthecaseofeucalyptol,trans-␤-ocimene,borneol, linalylacetate,lavandulylacetate,thePLSregressionmodelswere builtfromderivedNIRspectra,andforlinalool,camphor,and␤- caryophyllenefrombaselineoffsetpretreatedspectra.Toclassify samples,inthecaseoftheFIvariety,thePLS-DAregressionmodel wasbuiltfromderivedNIRspectraand forothervarietiesfrom baselineoffsetpretreatedspectra.

2.5.3. Software

ChemometricanalysiswasperformedusingTheUNSCRAMBLER XV.10.3(CAMO/Software,Oslo,Norway).

3. Resultsanddiscussion

3.1. Gaschromatography

Theidentificationofthecompoundswasachievedbycomparing theirmassspectrawiththoseofWiley275andNIST05alibrariesas

Fig.2. Scoreplot(PC1/PC3)ofthePCAofNIRdata(n=160).LavenderEOs:Fine(FI),Maillette(MA)andMatherone(MT)andlavandinEOs:Abrial(AB),Grosso(GR),Super (SP)andSumian(SU).

Fig.3.SuperpositionofthefirstregressionvectorsobtainedforFIvarietyandlavandulylacetate(a)SuperpositionofthefirstregressionvectorsobtainedforMAvarietyand eucalyptol(b)SuperpositionofthefirstregressionvectorsobtainedforMTvarietyandlinalool(c).

wellasbycomparingtheirretentionindiceswiththoseofauthen- ticsamples.Tables1and2listtherelativecomposition(meanand range)determinedfrompeakareasforthe28majorcompounds, eachaccountingformorethan0.1%intheEOs.Concerninglavender EOs,themajorcompoundsidentifiedarelinalylacetate(39.40%), linalool (28.58%), ␤-caryophyllene (4.96%), eucalyptol coeluted withcis-␤-ocimene(3.84%),trans-␤-ocimene(3.80%)andlavandu-

lylacetate(3.61%).ConcerninglavandinEOs,themajorcompounds identifiedarelinalool(36.74%),linalylacetate(29.08%),camphor (6.71%),eucalyptolcoelutedwithcis-␤-ocimene(6.18%)andbor- neol(3.80%).Formostsamples,lavenderEOswerecharacterizedby ahighercontentoflinalylacetatethanlinalool,whilelavandinEOs presentedhigheramountsoflinaloolthanlinalylacetate,exceptfor theSPvariety.AsshowninTables1and2,thereareothersignif-

Fig.4. SuperpositionofthefirstregressionvectorsobtainedforABvarietyandeucalyptol(a).SuperpositionofthefirstregressionvectorsobtainedforGRvarietyand trans-␤-ocimene(b).SuperpositionofthefirstregressionvectorsobtainedforSPvarietyand␤-caryophyllene(c).Superpositionofthefirstregressionvectorsobtainedfor SUvarietyandlinalylacetate(d).

icantdifferencesbetweenlavenderandlavandinEOs,particularly forcamphorand␤-caryophyllenecontents.Thelowcamphorcon- tentinlavender(0.39%)comparedwithlavandin(6.71%)justifiesits useintheperfumeindustryandtheuseoflavandininthetoiletries industry.The high␤-caryophyllenecontent inlavender (4.96%) comparedwithlavandin(1.79%)alsocharacterizeslavenders.Other minordifferences incompound contents wereobserved,show- ingapotentialdifferentiationbetweenlavenderandlavandinEOs.

Regarding lavender varieties, thelow linalool content (19.74%) andhightrans-␤-ocimenecontent(7.42%)inMTarecharacteris- ticsofthis variety.Theloweucalyptolcontent(1.43%) andhigh 3-octanonecontent(1.37%)inMAarecharacteristicsofthisvari- ety.Finevarietyischaracterizedbyhighcontentinterpinen-4-ol (4.12%).Regardinglavandinvarieties,thelowlinalylacetatecon- tentinSU(21.72%)andthehigheucalyptolcontentinAB(8.41%)are characteristicofthesevarieties.Thelowtrans-␤-ocimenecontent (0.29%)ischaracteristicofGRvariety.Supervarietyischaracterized byhighlinalylacetatecontent(38.02%)amonglavandinEOsandis alsocharacterizedbylow␤-caryophyllenecontent(1.36%).

3.2. NIRspectroscopy

Fig.1showstheNIRspectraoflavenderandlavandinEOsamples overthespectralrange4500–9000cm⁻¹.Eventhoughthecompo- sitionsoflavenderandlavandinEOsdiffergreatly,theirspectraare onlyslightlydifferent.Themajordifferencesbetweenthespec- trawere:theintensityinthe4950–5400cm⁻¹regionwhichwas strongerforlavandinEOs(secondovertoneofC Ostretchingvibra- tions);theintensityforthebandsat8422cm⁻¹,5887cm⁻¹ and

5766cm⁻¹ whichwerestrongerforlavandinEOs;andalsoarel- evantdifferenceofintensityat5624cm⁻¹whichwasstrongerfor lavenderEOs.Thespectraoftheeightmajorpureterpenoids(>2%) inlavenderandlavandinEOswererecordedandtheassignmentof majorbandswascarriedoutusingliteraturedata(Cozzolinoetal., 2005;Dupuyetal.,2010;Guoetal.,2006;WorkmanandWeyer, 2007).Thebandsbetween4445and4895cm⁻¹werecharacteris- ticofthecombinationofCHstretchingvibrationsofCH₃andCH₂ withothervibrations(C OandC Cstretching);thebandsbetween 4950and 5400cm⁻¹ wereattributedtothesecondovertoneof carbonylcompounds;thebandsbetween5690and6190cm⁻¹cor- respondedtothefirstovertoneoftheCHstretchingvibrationof CH₃,CH₂,and CH CH;thebandsbetween6600and7600cm⁻¹ correspondedtothefirstovertonecombinationsoftheCHstretch- ingvibrationofCH₃,CH₂,andCH CHandthebandsbetween8100 and8950cm⁻¹ correspondedtothesecondoftheCHstretching vibrationsofCH3,CH2,andCH CH.

3.3. Chemometricanalysis

Thefirstandthirdcomponents(PC1/PC3)ofthePCAcarriedout onNIRdatarepresent78%ofthetotalspectralvariance(Fig.2).

PrincipalcomponentsoneandthreewerepresentedinsteadofPC1 andPC2becausePC1andPC2allowedfordiscriminationbetween lavenderandlavandinEOsbutnotfordiscriminationbetweenvari- eties.Thelavenderandlavandingroupsareseparatedonthefirst component(PC1),lavenderisnegativelyprojectedwhilelavandin ispositivelyprojected,exceptfortheSPvarietywhichisnegatively projectedbutnearthecenter.Principalcomponentonealsoallows

Table3

StatisticsofthePLSregressionmodelsforthe8maincompoundsinthelavenderandlavandinEOsestablishedfromNIRdata.

Compounds Mean(GC%) R² SEC LV Q² SEP REP(%)

Eucalyptol^a 5.01 0.987 0.379 8 0.983 0.443 8.84

trans-␤-Ocimene^a 2.70 0.977 0.435 5 0.988 0.360 13.33

Linalool^b 32.66 0.991 0.978 5 0.993 0.847 2.59

Camphor^b 3.55 0.998 0.185 6 0.998 0.226 6.36

Borneol^a 2.51 0.994 0.174 6 0.994 0.200 7.97

Linalylacetate^a 34.24 0.994 0.782 3 0.993 0.821 2.39

Lavandulylacetate^a 2.55 0.987 0.306 8 0.971 0.424 16.62

␤-Caryophyllene^b 3.37 0.989 0.253 7 0.986 0.284 8.42

aPredictionregressionmodel(PLSregression)establishedfromderivedNIRdata.

b Predictionregressionmodel(PLSregression)establishedfrombaselineoffsetNIRdata.R²:coefficientscorrelationincalibration,Q²:coefficientscorrelationinprediction, SEC:standarderrorofcalibration,SEP:standarderrororprediction,LV:latentvariablesandREP:relativeerrorofprediction.

fortheseparationoflavandinvarieties.Principalcomponentone canbedescribedwithlinalylacetateandcamphorcontents.Laven- derEOshavelowcamphorcontent(0.39%)andhighlinalylacetate (39.40%)comparedwithlavandin(6.71%and29.08%respectively).

Asobservedpreviously,amonglavandinEOs theSPvarietywas characterizedbyhighlinalylacetateandlowcamphorcontents.

ThatiswhythisvarietyisveryclosetolavenderEOsinthePCA.Prin- cipalcomponentthreeallowsforseparationoflavendervarieties andcanbedescribedbylinaloolandeucalyptolcontents.Thevari- etiesMA,FIandMTareseparatedonPC3accordingtotheirlinalool content(39.09%,26.92%and19.74%respectively)andeucalyptol content(1.43%,4.13%and5.95%respectively).

The determination of the main compound contents in the EOs by chemometric analysis from NIR spectra in the range 5500–6500cm⁻¹ (a part of first overtone information) was achieved using PLS regression algorithms using gas chromato- graphicdataasreference.Table3givesthestatisticsof thePLS regressionmodelsfortheeightquantifiedcompounds(accounting formorethan2%).Forbestresults,inthecaseofeucalyptol,trans-

␤-ocimene,borneol,linalylacetateandlavandulylacetate,thePLS regressionmodelswerebuiltfromderivedNIRspectraandforother compoundsfrombaselineoffsetpretreatedNIRspectra.Asshown inTable3,theeightmaincompoundscanbeclassifiedinthree groups.Forthefirstgroup,comprisingthetwomajorcompounds, verygoodresultswereobtainedwitha REPat2.39%and 2.59%

forthepredictionof linalylacetateand linaloolrespectively. In thesecondgroup,fourcompoundswerecorrectlypredictedsince theirREPwerebetween5and10%(eucalyptol,camphor,borneol and␤-caryophyllene).Inthethirdgroup,twocompoundswerenot well-predictedwithaREPat13.33%and16.62%fortheprediction oftrans-␤-ocimeneandlavandulylacetaterespectively.Theper- formanceofthePLSpredictionmodelswasdirectlyrelatedtothe compoundcontentsintheEOs.

Table4 givesthe predictionresultsfor the53 lavender and lavandinEOsrecognition.Themodelswerebasedonthesamecali- brationandpredictionsetusedforquantitativeanalysis.Inthefirst stepthediscriminationbetweenlavenderandlavandinwasstud- ied.Theresultswereexcellent,with100%ofcorrectclassification.

Inthesecondstep,thesevenvarietieswerediscriminatedandwell- predictedwitha percentageofcorrectclassificationhigherthan 96%(Table4)ThevarietiesMA,MTandSPgave100%ofcorrect classification.Forlavendervarieties,theFIvarietypresentedtwo falsenegativesamples.Regardingthelavandinvarieties,ABpre- sentedonefalsepositivesample,GRpresentedonefalsepositive sampletooandSUpresentedonefalsenegativesample.Topredict varietalorigin,thedifficultycomesfromthevariabilityinthequal- ityofsamples,whichdependsonharvestyear,collectarea,storage conditionsandextractionprocess,regardlessofvarietalorigin.

Asit waspossibletoclassifysamplesasafunctionoflaven- der/lavandinand theoriginof theirvarieties, itwasinteresting tounderstandhowtheseclassificationswereestablishedinorder to identify metabolomic indicators. Figs. 3 and 4 present the superpositionof the first regression coefficients B obtainedfor lavenderandlavandinvarietiesrespectivelyandtheircorrespond- ingmetabolomicindicators.Forthelavendervarieties,theFIfirst regressioncoefficientBwascorrelatedwiththeregressioncoef- ficient obtained for quantitative analysis of lavandulyl acetate (Fig. 3a). ThePearson coefficient betweenthe firsttwo regres- sioncoefficientswas0.86.Finevarietywascharacterizedbythe highercontentoflavandulylacetate(meanof4.60%).Lavandulyl acetatecouldbeconsideredas ametabolomic indicatorforthe FIvariety.SomeMatheronesampleshadhighconcentrationsin lavandulylacetateandthiscouldexplainthatthreesamplesofthe MTvarietyoverlaytheFIgrouponthePCAscoreplot.Maillette firstregressioncoefficientBwasanti-correlatedwiththeregres- sioncoefficient obtainedfor quantitative analysisof eucalyptol (Fig.3b).ThePearsoncoefficientbetweenthetwofirstregression coefficientswas−0.97.Maillettevarietywascharacterizedbyalow content(mean1.43%)ineucalyptol.Eucalyptolcouldbeconsid- eredasametabolomicindicatorfortheMAvariety.Matheronefirst regression coefficientBwasanti-correlatedwiththeregression coefficientobtainedforquantitativeanalysisoflinalool(Fig.3c).

ThePearsoncoefficientbetweenthefirsttwo regression coeffi- cientswas−0.99.Matheronevarietywascharacterizedbyalow content(19.74%)inlinalool,whichwasametabolomicindicator forthisvariety.Concerningthelavandinvarieties,theABandcam-

Table4

Classificationmatrixobtainedinprediction(PLS-DAregression)oflavenderandlavandinEOvarietiesestablishedfromNIRdata.

Variety Numberof latentvariables

AB (n=5)

GR (n=12)

SP (n=8)

SU (n=6)

FI (n=7)

MA (n=10)

MT (n=5)

False-negative samples

False-positive samples

NIRmodels Lavandin AB^a 12 5 0 0 1 0 0 0 0 1

GR^a 10 0 12 1 0 0 0 0 0 1

SP^a 10 0 0 8 0 0 0 0 0 0

SU^a 8 0 0 0 5 0 0 0 1 0

Lavender FI^b 8 0 0 0 0 5 0 0 2 0

MA^a 11 0 0 0 0 0 10 0 0 0

MT^a 8 0 0 0 0 0 0 5 0 0

aPredictionregressionmodel(PLSregression)establishedfrombaselineoffsetNIRdata.

b Predictionregressionmodel(PLSregression)establishedfromderivedNIRdata.