ArticlePDF Available

Classification of Smoke Contaminated Cabernet Sauvignon Berries and Leaves Based on Chemical Fingerprinting and Machine Learning Algorithms

Article

Classification of Smoke Contaminated Cabernet Sauvignon Berries and Leaves Based on Chemical Fingerprinting and Machine Learning Algorithms

Abstract and Figures

Wildfires are an increasing problem worldwide, with their number and intensity predicted to rise due to climate change. When fires occur close to vineyards, this can result in grapevine smoke contamination, and subsequently, the development of smoke taint in wine. Currently, there are no in-field detection systems that growers can use to assess whether their grapevines have been contaminated by smoke. This study evaluated the use of near-infrared (NIR) spectroscopy as a chemical fingerprinting tool, coupled with machine learning, to create a rapid, non-destructive in-field detection system for assessing grapevine smoke contamination. Two artificial neural network models were developed using grapevine leaf spectra (Model 1) and grape spectra (Model 2) as inputs, and smoke treatments as targets. Both models displayed high overall accuracies in classifying the spectral readings according to the smoking treatments (Model 1: 98.00%; Model 2: 97.40%). Ultraviolet to visible spectroscopy was also used to assess the physiological performance and senescence of leaves, and the degree of ripening and anthocyanin content of grapes. The results showed that chemical fingerprinting and machine learning might offer a rapid, in-field detection system for grapevine smoke contamination that will enable growers to make timely decisions following a bushfire event, e.g., avoiding harvest of heavily contaminated grapes for winemaking or assisting with a sample collection of grapes for chemical analysis of smoke taint markers.
Content may be subject to copyright.
Sensors2020,20,5099;doi:10.3390/s20185099www.mdpi.com/journal/sensors
Article
ClassificationofSmokeContaminatedCabernet
SauvignonBerriesandLeavesBasedonChemical
FingerprintingandMachineLearningAlgorithms
VasilikiSummerson1,ClaudiaGonzalezViejo1,ColleenSzeto2,3,KerryL.Wilkinson2,3,
DamirD.Torrico4,AlexisPang1,RobertaDeBei2andSigfredoFuentes1,*
1DigitalAgriculture,Food,andWineGroup,FacultyofVeterinaryandAgriculturalSciences,
TheUniversityofMelbourne,Parkville3010,Victoria,Australia;
vsummerson@student.unimelb.edu.au(V.S.);cgonzalez2@unimelb.edu.au(C.G.V.);
alexis.pang@unimelb.edu.au(A.P.)
2SchoolofAgriculture,FoodandWine,TheUniversityofAdelaide,WaiteCampus,PMB1,GlenOsmond,
SA5064,Australia;colleen.szeto@adelaide.edu.au(C.S.);kerry.wilkinson@adelaide.edu.au(K.L.W.);
roberta.debei@adelaide.edu.au(R.D.B.)
3TheAustralianResearchCouncilTrainingCentreforInnovativeWineProduction,PMB1,GlenOsmond,
SA5064,Australia
4DepartmentofWine,FoodandMolecularBiosciences,FacultyofAgricultureandLifeSciences,Lincoln
University,Lincoln7647,Canterbury,NewZealand;Damir.Torrico@lincoln.ac.nz
*Correspondence:sfuentes@unimelb.edu.au
Received:22August2020;Accepted:5September2020;Published:7September2020
Abstract:Wildfiresareanincreasingproblemworldwide,withtheirnumberandintensity
predictedtoriseduetoclimatechange.Whenfiresoccurclosetovineyards,thiscanresultin
grapevinesmokecontaminationand,subsequently,thedevelopmentofsmoketaintinwine.
Currently,therearenoinfielddetectionsystemsthatgrowerscanusetoassesswhethertheir
grapevineshavebeencontaminatedbysmoke.Thisstudyevaluatedtheuseofnearinfrared(NIR)
spectroscopyasachemicalfingerprintingtool,coupledwithmachinelearning,tocreatearapid,
nondestructiveinfielddetectionsystemforassessinggrapevinesmokecontamination.Two
artificialneuralnetworkmodelsweredevelopedusinggrapevineleafspectra(Model1)andgrape
spectra(Model2)asinputs,andsmoketreatmentsastargets.Bothmodelsdisplayedhighoverall
accuraciesinclassifyingthespectralreadingsaccordingtothesmokingtreatments(Model1:
98.00%;Model2:97.40%).Ultraviolettovisiblespectroscopywasalsousedtoassessthe
physiologicalperformanceandsenescenceofleaves,andthedegreeofripeningandanthocyanin
contentofgrapes.Theresultsshowedthatchemicalfingerprintingandmachinelearningmight
offerarapid,infielddetectionsystemforgrapevinesmokecontaminationthatwillenablegrowers
tomaketimelydecisionsfollowingabushfireevent,e.g.,avoidingharvestofheavilycontaminated
grapesforwinemakingorassistingwithasamplecollectionofgrapesforchemicalanalysisof
smoketaintmarkers.
Keywords:smoketaint;remotesensing;climatechange;nearinfraredspectroscopy;volatile
phenols
1.Introduction
Theincidenceandintensityofwildfiresareincreasingworldwide,mainlyduetotheeffectsof
climatechange[1–5].Bushfiresthatoccurnearwineregionscanresultingrapevinesmokeexposure,
whichcanalterthechemicalcompositionofgrapeberries.Wineproducedfromthesesmokeaffected
grapesmayexhibitunpalatablesmokyaromasandflavors,suchas“burntwood”,“ashy”,and“burnt
Sensors2020,20,50992of24
rubber”[6–9].Theseundesirablecharactershavebeenattributedtosmokederivedvolatilephenols
(VPs),includingguaiacol,4methylguaiacol,cresols,andsyringol[7,10,11].Itisthoughtthatthese
VPsaccumulateprimarilyintheskinofgrapeberriesfollowingsmokeexposureand,toalesser
extent,inthepulpandseeds[12–15].Grapevinesmokeexposure,andtheresultingsmoketaintin
wine,havecausedsignificantfinanciallossesforgrapegrowersandwinemakersduetodiscarded
grapesandunsaleablewine.Forexample,the2009BlackSaturdaybushfiresinVictoria,Australia,
wereestimatedtohavecausedAUD300millioninlostrevenue[16–19].Morerecently,theAustralian
GrapeandWineIncorporated(AWGI)estimatedanAUD40millionlossfromthe2019/2020summer
bushfires[20].Vineyardsmokeexposure,therefore,remainsasignificantissueforthewineindustry,
particularlygiventheincreasingfrequencyandseverityofbushfires[21].
GrapevineleaveshavealsobeenfoundtoaccumulateVPs,andapositivecorrelationhasbeen
demonstratedbetweenthelevelsofsmokecompoundsdetectedinleavesandwinewhentheywere
includedintheprimaryfermentation[13,22,23].Fromaphysiologicalpointofview,smokeexposure
hasalsobeenshowntodecreasestomatalconductanceinleaves,whichmayresultfromthereaction
ofcarbondioxide(CO2)andcarbonmonoxide(CO)withwatervaporinthesubstomatalcavity
producingcarbonicacid(H2CO3)[24,25].CarbonicacidreducesthepHinthestomata,resultingin
partialorcompletestomatalclosure[25,26].Damagetoleafsurfacesfollowingsmokeexposurehas
alsobeenobserved,withthedevelopmentofnecroticlesionsor,inextremecases,totalleafnecrosis
[10,22,27].Thismaybetheresultofozone(O3)presentinsmoke,whichhasbeenlinkedtochlorophyll
destructionandacceleratedleafsenescence[28,29].
Somechromatographictechniquessuchasgaschromatographymassspectrometry(GCMS)
andhighperformanceliquidchromatographytandemmassspectrometry(HPLCMS/MS)have
beendevelopedtoquantifylevelsoffreeandglycosidicallyboundVPsingrapesandwines[30–33].
Whilethesetechniquesarecurrentlyusedforqualitativeandquantitativeanalysisandmayassist
growersindeterminingthelevelofsmoketaintinthefinalwine,therearenumerousshortcomings:
samplepreparationistimeconsuminganddestructive,andanalysesrequireexpensivereagents,
standards,andequipment,aswellastrainedpersonnel.Furthermore,followingabushfireevent,
theremaybelongdelaysintheavailabilityofresultsduetolargenumbersofsamplesbeing
submittedtocommerciallaboratoriesforanalysis[34,35].Consequently,alternativemethodsof
smoketaintanalysishaverecentlybeeninvestigatedandmayoffernondestructivesample
preparation,aswellasaccurateandrapidresults.
Theuseofspectroscopictechniqueshasincreasedinrecentyearsduetotheireaseofuse,rapid
results,minimalsamplepreparation,andnondestructivenature,allofwhichallowrepeated
measurementstobetaken[34–39].Furthermore,thedevelopmentofsmaller,handheldspectroscopic
devicescoupledwithdecreasingcosts,hasallowedthesetechnologiestobemorereadilyaccessible
andaffordabletogrowersandfarmers,whiletheirportabilityallowsforinfielduse,reducingthe
riskofsampledeteriorationduringtransportation[39,40].Ultraviolet(UV)tovisible(Vis)
spectroscopyinvolvestheregionbetween200–780nm,whichcanbeusedtoanalyzecompounds
containingorganicacids,phenoliccompounds,andpigmentssuchasanthocyanins,carotenoids,and
chlorophylls[41].UVVisspectroscopyhasbeenusedtodeterminethecontributionofchemical
compoundstowardsthecompositionofextravirginoliveoilstodeterminetheregioninthe
Mediterraneanitwasproduced,tooptimizetheagingprocessofSpanishwines,andtoassessthe
impactofheatingedibleoilsandtodeterminetheiracidlevel[42–44].Nearinfrared(NIR)
spectroscopybetweenthelightspectraregionsof780–2500nmhasbeenwidelyusedinagricultural
andfoodscienceapplications,withNIRbandscorrespondingtoovertonesresultingfromthe
vibrationsofOH,CH,NH,andSHbonds[39,41].Variousspectroscopictechniques,mostnotably
intheNIRregion,havebeenusedfornumerousapplicationsinviticulture,includingtheassessment
ofgrapequalityandripenessaswellastheauthenticationofgeographicalorigin[38,45–50].Research
hasalsobeenconductedontheuseofmidinfrared(MIR)spectroscopy(between2500–25,000nm)of
theelectromagneticspectrum,aswellassynchronoustwodimensionalMIRcorrelationspectroscopy
(2DCOS)fortheclassificationofsmoketaintedwines[34,35].Bothtechniquesshowedpotentialfor
screeningsmoketaintedwine,withMIRspectroscopyachieving61and70%classificationratesfor
Sensors2020,20,50993of24
controlandsmokeaffectedwines,respectively.However,classificationrateswereaffectedbythe
degreeofsmoketaint,aswellascompositionaldifferencesarisingfromthegrapevarietyandoak
maturation[34].Whilethistechnologymayhelptoassesswinesamplesforsmoketaint,itdoesnot
provideanearly,infielddetectionsystemthatcouldhelpgrowersidentifywhichgrapesmaybe
contaminatedbeforewinemaking.Atpresent,thereisverylittleresearchinvestigatingtheinfield
useofVisNIRspectroscopyfortheclassificationofsmokeaffectedgrapevineleavesandberries.
ResearchbyFuentesandcoworkers[19]developedamodelusingNIRspectroscopybetweenthe
regionof700–1100nmtopredictthelevelsofguaiacolglycoconjugatesinberriesandwine,andthe
levelsofguaiacolinwine.Thesemodelsmayoffergrowersanondestructiveinfielddetection
systemforgrapevinesmokecontamination.However,furtherresearchisrequiredtodeterminethe
effectivenessofdifferentNIRregionsformonitoringsmokecontamination.
Severalchemometrictechniqueshavebeenusedtoanalyzespectraldata,includingpartialleast
squares(PLS)regression,principalcomponentanalysis(PCA),andartificialneuralnetworks(ANN),
tonameafew[41].Ofthesetechniques,ANNshaveincreasedinpopularityasclassification,
prediction,andclusteringtools,particularlysincetheycanbetterinterpretthenonlinearpatternsof
spectraldata[51–54].Machinelearning(ML)modelingbasedonANNcanbetrainedfromasetof
givendataknownas‘inputs’orindependentvariablesandformcomplex,nonlinearrelationships
withtheseinputsandthe‘targets’ordependentvariables[54].Forexample,preliminaryMLmodels
fortheclassificationofsmoketaintedgrapevineshavebeendevelopedusinginfrared(IR)thermal
imageryfromcanopies,whichgaveanindicationofchangesinstomatalconductancefor
classificationofcontrolandsmokeexposedgrapevines[25].Inadditiontothis,anothermodelhas
beenproposedthataimstoquantifylevelsofsmokederivedcompoundsingrapesandwineusing
NIRspectroscopymeasurementsasinputs[25].Furthermore,UVVisspectroscopymayoffer
insightsintothedegreeofphysiologicalperformanceofleavesaswellasfruitripeningandquality
throughanalyzingpigmentcontent,suchaschlorophylls,anthocyanins,andcarotenoids[55–59].
TheobjectiveofthisstudywastoinvestigatetheuseofNIRspectroscopy,coupledwithML
modelingforthedetectionofgrapevinesmokecontamination.Grapevineleavesandberrieswere
analyzedinthevineyardinasmoketrialusingaNIRspectrometer,andtheabsorbancevalueswere
usedasinputstotraindifferentmachinelearningalgorithmsinordertocreateANNswiththebest
classificationperformances.Inadditiontothis,UVVisspectroscopywasusedtoassessthe
physiologicalperformanceanddegreeofsenescenceofleaves,aswellasthedegreeofripeningand
anthocyanincontentofgrapes.Thismayoffergrowersarapidandnondestructivedetectionsystem
thattheycanemploythemselvestoobtainrealtimeinformationregardingsmokeexposure.Thiswill
facilitatetimelydecisionmakingaroundwhichfruittosampleforchemicalanalysisand/orto
harvesttomaintainwinequality.
2.MaterialsandMethods
2.1.VineyardSiteandExperimentalDesignfortheSmokeTrial
ThesmoketrialwasconductedinlateJanuaryearlyFebruaryduringthe2018/2019growing
season,attheUniversityofAdelaide’sWaiteCampusinUrrbrae,SouthAustralia(34°58′S,138°38′
E).Thetrial,describedpreviouslybySzetoandcolleagues[60],involvedtheapplicationofsmoke
and/orincanopymistingtoCabernetSauvignongrapevinesandcomprisedfivedifferenttreatments:
acontrol(C),i.e.,neithermistingnorsmokeexposure;(ii)acontrolwithmisting(CM),i.e.,incanopy
mistingbutnosmokeexposure;(iii)ahighdensitysmoketreatment(HS);(iv)ahighdensitysmoke
treatmentwithmisting(HSM);and(v)alowdensitysmoketreatmentwithoutmisting(LS).
TreatmentswereappliedtoCabernetSauvignongrapevinesplantedin1998at2.0and3.3mvineand
rowspacings,andtrainedtoabilateralcordon,verticalshootpositionedtrellissystem(VSP),hand
prunedtoatwonodespursystem,withundervinedripirrigation(twiceweekly,fromfruitsetto
preharvest).Smoketreatmentswereapplied(approximatelysevendayspostvéraison,theperiod
grapesarethoughttobemostsusceptibletosmokecontamination[10])usingapurposebuiltsmoke
tent(Figure1a,b)andexperimentalconditionsreportedpreviously[4,61]:lowandhighdensity
Sensors2020,20,50994of24
smoketreatmentswereachievedbyburningdifferentfuelloads(i.e.,~1.5and5kgofbarleystraw,
respectively).Incanopymistingwasevaluatedasamethodformitigatingtheuptakeofsmoke
derivedvolatilephenolsbygrapesandinvolvedthecontinuousapplicationoffinewaterdroplets
(65μm)tothegrapevinebunchzoneusingapurposebuiltsprinklersystem(deliveringwaterat11
L/h),aspreviouslydescribed[62].Eachtreatmentwasappliedtosixvinesfromthreeadjacentpanels,
excepttheHStreatment,whichcomprisedonlyfivevines,withtreatmentsseparatedbyatleastone
buffervine.LS,HS,andHSMtreatmentscomprisedduplicateapplicationsofsmoketo1.5
panels/threevinesatatime(exceptforoneHStreatment).Theincanopysprinklersystemwasturned
on5minbeforethefirstHSMtreatmentwasappliedandoff15minafterthesecondHSMtreatment
wascompleted,suchthatCMandHSMgrapevinesweremistedforapproximately2.5hintotal.The
secondandfifthvinefromeachtreatment(themiddlevinesfromsmoketreatments)werethen
selectedforphysiologicalandNIRmeasurements.
(a)(b)
Figure1.Smoketreatmentswereappliedtograpevinesusingapurposebuiltsmoketent;grapevines
wereenclosedinthetentandexposedtosmokederivedfromthecombustionofbarleystraw(a,b).
2.2.PhysiologicalMeasurements
Therateofphotosynthesis(A),stomatalconductance(gs),andtranspiration(E)weredetermined
usingaportableinfraredgasanalyzerequippedwithabroadleafchamber(LCproSD,ADC
BioscientificLtd.,Hoddesdon,UK).Measurementsweretakenonthreeleavesofeachsideofthe
canopypervine(n=12leavespertreatment)withaphotosyntheticphotonfluxdensityof1000μmol
m2s1suppliedbyahighefficiency,lowheatoutput,mixedredbluelightemittingdiode(LED)
arrayunit.WatervaporandCO2concentrationinthechamberweresettoambient.Measurements
weretakenoneday(24h)aftersmoketreatmentswereapplied,onclear,sunnydays.
2.3.DeterminationofVolatilePhenolsandTheirGlycoconjugatesinGrapeJuice/Homogenate
Theconcentrationofvolatilephenolsandtheirglycoconjugatesweredetermined(ingrapejuice
andhomogenate,respectively)usinganalyticalmethodsdescribedpreviously[30,32,33,60].Volatile
phenolsweremeasuredbystableisotopedilutionanalysis(SIDA)[3,30,33],usinganAgilent6890
gaschromatographcoupledtoa5973massspectrometer(AgilentTechnologies,ForestHill,Vic.,
Australia).Isotopicallylabeledstandards,i.e.,d4guaiacolandd3syringol,werepreparedinhouse
Sensors2020,20,50995of24
usingmethodsoutlinedpreviously[3,30,33].Thelimitofquantitationforvolatilephenolswas1–2
μg/L.VolatilephenolglycoconjugateswerealsomeasuredbySIDA[30,32],usinganAgilent1200
highperformanceliquidchromatograph(HPLC)equippedwitha1290binarypump,coupledtoan
ABSCIEXTripleQuadTM4500tandemmassspectrometer,withaTurboVTMionsource(Framingham,
MA,USA).Thepreparationoftheisotopicallylabeledinternalstandard,i.e.,d3syringol
gentiobioside,hasbeenreportedpreviously[30,32].Thelimitofquantitationforvolatilephenol
glycosideswas1μg/kg.
2.4.NearInfraredDataCollection
Grapevineleafandberryspectrawerecollectedonedayaftersmokeexposure,usinga
microPHAZIRTMRXAnalyzer(ThermoFisherScientific,Waltham,MA,USA),whichhadaspectral
rangeof1596to2396nmatintervalsof7–9nm.Priortoundertakingthemeasurementsandafter
every10–15readings,thedevicewascalibratedusingawhitebackgroundcalibrationstandard
(includedwiththedevice).Thewhitebackgroundwasplacedontopoftheleafwhilemeasuringto
avoidsignalnoiseinclusionduetovariationinlightorenvironmentalchanges.Leavesandberries
werealsoanalyzedusingtheLightingPassportProTMhandheldspectrometer(Asensetek
Incorporation,XindianDistrict,NewTaipeiCity,Taiwan),whichhasaspectralrangeof380–780nm
atintervalsof1nm.Measurementsweretakenatapproximately3cmfromtheleavesandberries.
Allmeasurementswereconductedatambienttemperaturebetween9:00a.m.and6:00p.m.
Fortheleafspectralmeasurements,ninesunlitandnineshaded,mature,fullyexpandedleaves
wereselected(i.e.,18leavespervine,36leavespertreatment).Leaveswerefreeofanyvisiblesigns
ofdiseaseorblemishes.Eachleafwasmeasuredinthreeareas,intriplicate,usingthe
microPHAZIRTMRXAnalyzer,whilethreemeasurementsperleafweretakenwiththeLighting
PassportProTMhandheldspectrometer.Fortheberryspectra,twobuncheswereselectedpervine,
andnineberries(threefromthetop,middle,andbottomofeachbunch)weremeasured,intriplicates
usingthemicroPHAZIRTMRXAnalyzer(n=540).Ontheotherhand,twelveberriespertreatment
wereanalyzedusingtheLightingPassportProTM(n=180)whilestillattachedtothebunch.
2.5.CalculatingSpectralIndices
Spectralindicesfortheanalysisofpigmentcontentwerecalculatedforbothleavesandberries.
LeafspectratakenusingtheLightingPassportProTMwereusedtocalculatethenormalizeddifference
vegetationindex(NDVI),normalizedanthocyaninindex(NAI),plantsenescencereflectanceindex
(PSRI),andcarotenoidreflectanceindex(CRI)[56,57,59,63–65].Berryspectrawereusedtocalculate
theNAIandPSRI.Thecalculationsandwavelengthsusedfordeterminingtheseindicesaregivenin
Table1.
Table1.Calculationsforthespectralindicesinvestigatedinthisstudy.
IndexNameIndexAbbreviationEquationReferences
NormalizeddifferencevegetationindexNDVI


780 660
780 + 660
II
II
[56,57]
NormalizedanthocyaninindexNAI

0
()780 570
780+ 57
II
II
[56,57]
CarotenoidreflectanceindexCRI5500
11
50 515II
[63,64]
CarotenoidreflectanceindexCRI7000
11
50 710II
[65]
PlantsenescencereflectanceindexPSRI680 500
750
II
I
[59]
Sensors2020,20,50996of24
2.6.StatisticalAnalysis
Physiologicalmeasurements,spectralindices,volatilephenols,andtheirglycoconjugateswere
analyzedbyonewayanalysisofvariance(ANOVA)usingMinitab®version18.1(MinitabInc.,State
College,PA,USA).MeancomparisonswereperformedusingtheFisherleastsignificantdifference
(LSD)methodasaposthoctestatα=0.05.NearinfrareddatawereanalyzedusingTheUnscrambler
Xversion10.3software(CAMOSoftware,Oslo,Norway).Absorbancevaluesforallwavelengths
wereplottedforboththemicroPHAZIR
TM
RXAnalyzerandLightingPassportPro
TM
leafandberry
readings.Principalcomponentanalysis(PCA)wasalsoperformedusingTheUnscramblerX
program.AllmicroPHAZIR
TM
RXAnalyzermeasurementswerepreprocessedusingthesecond
derivativetransformation,Savitzky–Golayderivation,andsmoothingusingTheUnscramblerX
version10.3softwarepriortotheplottingofgraphsandstatisticalanalysis.
2.7.ArtificialNeuralNetworkModeling
ThreeANNmodelsweredevelopedforberryandleafNIRreadings,whichwereusedasinputs
toclassifythedifferentsmoketreatmentsusingcustomizedcodewritteninMATLAB®(version
R2020a,MathWorksInc.,Natick,MA.USA)(Figure2).Thiscodetestedatotalof17training
algorithmsinalooptofindtheoptimumintermsofaccuracyandperformance.Oncetheoptimum
trainingalgorithmwasidentified,furthertrainingwasperformedtodevelopthemostaccurateANN
model.Forbothmodels,theLevenberg–Marquardttrainingalgorithmwasfoundtobethebest
algorithm,resultinginmodelswiththehighestaccuracyandnosignsofoverfitting.
Overtoneswithinthe1596–1800nmrangewereusedasinputsforthemicroPHAZIR
TM
leaf
model(Model1).Thisregionwasselectedtoavoidwaterovertonesandanyclassificationresulting
fromthewaterstatusofthevines.TheentirespectralrangewasusedforthemicroPHAZIR
TM
berry
model(Model2)(1596–2396nm).Thetwomodelsweredevelopedusingarandomdatadivisionwith
70%(n=1134forModel1and378forModel2)training,15%(n=243forModel1and81forModel
2)forvalidationwithameansquarederror(MSE)performancealgorithmand15%(n=243forModel
1and81forModel2)fortestingwithadefaultderivativefunction.Tenhiddenneuronswereselected
foreachofthetwomodelsafterconductingatrimmingexercisewiththree,five,andtenneurons.
(a)
(b)
Sensors2020,20,50997of24
Figure2.Twolayerfeedforwardnetworkwithtenhiddenneuronsandsigmoidfunctionforthethree
classificationmodels:(a).microPHAZIRTMleafmodel(Model1)and(b).microPHAZIRTMberry
model(Model2).Abbreviations:C=controlwithoutmisting;CM=controlwithmisting;HS=high
densitysmokewithoutmisting;HSM=highdensitysmokewithmisting;andLS=lowdensitysmoke.
3.Results
3.1.PhysiologicalMeasurements
ResultsofgasexchangeparametersareshowninTable2.Thetranspirationratewaslowerfor
theHStreatment(P<0.005)withameanrateof1.43mmolm2s1,whilenodifferenceswereobserved
intheothertreatments.TheCMandCtreatmentsbothhadthehighestgsvalueswithanaverage
valueof0.15molm2s1foreach,whileHSandLStreatmentshadthelowestaveragegsat0.056mol
m2s1and0.082molm2s1respectively.MeanratesofAwerefoundtobehighestintheCandCM
treatments(10.77μmolm2s1and9.66μmolm2s1,respectively),whiletheLSandHStreatments
hadthelowest(7.01μmolm2s1and5.59μmolm2s1,respectively).
Table2.Gasexchangeparametersmeasuredforthedifferentsmoketreatments.
SmokeTreatmentE(mmolm2s1)gs(molm2s1)A(μmolm2s1)
MeanSDMeanSDMeanSD
C2.48a0.700.15a0.0510.77a3.46
CM2.31a0.540.15a0.059.66ab2.31
HS1.43b0.620.06c0.035.59d2.8
HSM2.06a0.440.10b0.038.15bc1.97
LS2.18a0.780.08bc0.037.01cd2.42
Abbreviations:C=controlwithoutmisting;CM=controlwithmisting;HS=highdensitysmoke
withoutmisting;HSM=highdensitysmokewithmisting;andLS=lowdensitysmoke;SD=standard
deviation.MeansfollowedbydifferentlettersaresignificantlydifferentbasedonFisherleast
significantdifference(LSD)posthoctest(α=0.05).
3.2.LevelsofSmokeTaintMarkerCompoundsinGrapeJuice/Homogenate
DifferencesinvolatilephenolconcentrationsbetweenHSandHSMtreatmentswerefoundfor
guaiacol,4methylsyringol,andsyringol(P<0.05;TableS1).Inparticular,4methylsyringoland
syringolhadthelargestdifferencesinconcentrationsamongstthesmoketreatments,withtheHS
treatmentexhibitingthehighestmeanvalues(17and126μg/L,respectively)followedbytheHSM
treatment(9and59μg/L,respectively)whiletheCMtreatmentsexhibitedthelowestmeanvalues(2
and8μg/L),whichdisplayedthelowestmeanvalue.TherewerenodifferencesbetweentheHSand
HSMtreatments,norbetweentheC,CM,andLStreatmentsfor4methylguaiacol,phenol,andtotal
cresols;however,HSandHSMgrapeshadsignificantlyhighervolatilephenolconcentrationsthan
C,CM,andLSgrapes.
Somedifferencesinvolatilephenolglycoconjugatelevelscouldbeseenamongstthefivesmoke
treatments.SomeglycoconjugatesdisplayeddifferencesbetweentheHSandHSMtreatments.There
wasnodifferenceinGuRGlevelsbetweentheLS,HS,andHSMtreatments,withnolevelsdetected
intheCandCMtreatments.TheHSsmoketreatmenthadthehighestlevelsofPhRG,PhGG,CrPG,
SyGG,andSyPG,followedbytheHSMandLStreatmentsandthentheCandCMtreatments.
InterestinglytheCandHStreatmentshadthehighestlevelofCrGGfollowedbytheCMandHSM
treatment,whiletheLStreatmenthadthelowestconcentration.
3.3.NIRAbsorbancePatternsforLeavesandBerries
Absorbancespectrafortheaveragesofreplicatesforbothrawandtransformedleafabsorbance
spectraaredepictedinFigures3and4.ForthemicroPHAZIRTMRXAnalyzerleafabsorbances,clear
differencesinspectralreadingswereobservedforeachsmokingtreatment.Apeakwasobservedat
Sensors2020,20,50998of24
approximately1784–1793nm(Figure3a),whileforthetransformeddata(Figure3b),largepeaksare
presentbetween1596–1647nm.
(a)
(b)
Sensors2020,20,50999of24
Figure3.Rawleafabsorbance(a)andsecondderivativespectra(b)measuredwiththe
microPHAZIRTMnearinfrared(NIR)analyzerforthedifferentsmokeandmistingtreatments.
Abbreviations:C=controlwithoutmisting;CM=controlwithmisting;HS=highdensitysmoke
withoutmisting;HSM=highdensitysmokewithmisting;andLS=lowdensitysmoke.
DifferencesinabsorptionreadingswerealsofoundforthemicroPHAZIRTMRXAnalyzerberry
absorbancespectra(Figure4a).Peakswereoriginallyobservedatapproximately1785and1902nm,
butinthetransformeddata(Figure4b),largepeakswereobservedbetweenapproximately1596–
1640nmand1820–1940nm.
(a)
Sensors2020,20,509910of24
(b)
Figure4.Rawberryabsorbance(a)andsecondderivativespectra(b)measuredwiththe
microPHAZIRTMNIRanalyzerforthedifferentsmokeandmistingtreatments.Abbreviations:C=
controlwithoutmisting;CM=controlwithmisting;HS=highdensitysmokewithoutmisting;HSM
=highdensitysmokewithmisting;andLS=lowdensitysmoke.
3.4.PrincipalComponentAnalysis
Figure5ashowstheprincipalcomponentanalysis(PCA)forthemicroPHAZIRTMRXAnalyzer
leafspectrawithabsorbancevaluesbetween1600–1800nm.Thefirstprincipalcomponent(PC1)
accountedfor62%ofthedatavariability,whileprincipalcomponenttwo(PC2)accountedfor24%.
Hence,86%ofthetotalvariabilitywasexplainedbythesePCs.Therewasnoclearseparationofthe
differentsmoketreatmentswhenmodeledwiththemicroPHAZIRTMleafspectra.PC1was
representedbywavelengthsbetween1604–1621nmandbetween1621–1647nm(loadingsshownin
Figure5b).PC2wasrepresentedbywavelengthsbetween1613–1647nm,aswellas1604nm.
Sensors2020,20,509911of24
(a)
(b)
Figure5.Principalcomponentanalysis(PCA)forthemicroPHAZIR
TM
leafabsorbancevalues
between1600–1800nm(a)andloadings(b).Abbreviations:C=controlwithoutmisting;CM=control
withmisting;HS=highdensitysmokewithoutmisting;HSM=highdensitysmokewithmisting;and
LS=lowdensitysmoke.
Figure6ashowsthePCAforthemicroPHAZIR
TM
RXAnalyzerberryspectra,where59%ofthe
datavariabilitywasdescribedbyPC1,whilePC2accountedfor10%ofthedatavariability;thus,a
totalof69%ofthetotaldatavariabilitywasexplainedbythefirsttwocomponentsofthePCA.As
withthemicroPHAZIR
TM
RXAnalyzerleafspectra,mostofthesmoketreatmentsoverlapped
quadrants.TheCMtreatmentwasgroupedprimarilyintheupperrightquadrant,whileCandLS
treatmentsweregroupedprimarilyinthelowerright.TheHStreatmentwaslocatedprimarilyinthe
upperrightandleftquadrants,whiletheHSMtreatmentwasgroupedintheleftupperandlower
quadrants.PC1onewasrepresentedbythewavelengthregion1604–1622.PC2wasrepresentedby
thewavelengthsbetween1630–1647nmand2374–2389nm(loadingsshowninFigure6b).
Sensors2020,20,509912of24
(a)
(b)
Figure6.Principalcomponentanalysis(PCA)forthemicroPHAZIR
TM
berryabsorbancevalues
between1600–2396nm(a)andloadings(b).Abbreviations:C=controlwithoutmisting;CM=control
withmisting;HS=highdensitysmokewithoutmisting;HSM=highdensitysmokewithmisting;and
LS=lowdensitysmoke.
3.5.SpectralIndices
ResultsforthespectralindicesareshowninTable3.InthecaseoftheleafNDVIandNAI,the
HSandCtreatmentshadthelowestmeanvalues(0.72and0.64fortheHStreatmentand0.84and
0.74fortheC)(P<0.05).Therewerenodifferencesfortheremainingtreatments.FortheleafPSRI,
theHStreatmenthadthehighestmeanvalueat0.065,withnodifferencesfortheremaining
treatments.FortheleafCRI
500
,theLSandHStreatmentshadthehighestvaluesat1.45and1.20,
respectively,
andfortheCRI
700,
theLStreatmentshadthehighestmeanvaluesat1.76,withno
differencesfortheremainingtreatments.
Sensors2020,20,509913of24
InthecaseoftheberryNAI,theHSandLStreatmentshadthehighestmeanvalueswith0.88
and0.87,withboththeCandLStreatmentshavingthelowestmeanvaluesof0.80and0.75.Forthe
PSRI,boththeLSandCtreatmentshadthehighestmeanvaluesof0.02,whiletheHSMhadthe
lowestvalueat‐0.02.
Sensors2020,20,509914of24
Sensors2020,20,5099;doi:10.3390/s20185099www.mdpi.com/journal/sensors
Table3.Meansandstandarddeviation(SD)ofspectralindicescalculatedforleavesandberries.
Treatment
LeafBerry
NDVINAIPSRICRI500CRI700NAIPSRI
MeanSDMeanSDMeanSDMeanSDMeanSDMeanSDMeanSD
CM0.85a0.100.77a0.110.00b0.010.70b0.640.82b0.79‐ ‐
C0.84ab0.0820.74ab0.110.01b0.020.67b0.780.77b0.870.80b0.070.02a0.02
HS 0.72b0.500.64b0.490.07a0.191.20a0.240.82b0.620.88a0.040.00b0.00
HSM0.87a0.110.79a0.110.00b0.020.48b0.060.58b0.450.75b0.10−0.02c0.00
LS0.92a0.040.84a0.080.00b0.011.45a1.081.76a1.400.87a0.050.02a0.01
Abbreviations:C=controlwithoutmisting;CM=controlwithmisting;HS=highdensitysmokewithoutmisting;HSM=highdensitysmokewithmisting;andLS=low
densitysmoke.MeansfollowedbydifferentlettersarestatisticallysignificantbasedonFisher’sleastsignificantdifference(LSD)posthoctest(α=0.05).
Sensors2020,20,509915of24
3.6.ArtificialNeuralNetworkModels.
Table4showstheconfusionmatricesforthetwomodelsdevelopedusingthespectralreadings
asinputsandtheexperimentaltreatmentsastargets.Bothmodelsdisplayedhighaccuracyin
classifyingthespectralreadingsaccordingtothetreatments,withanoverallaccuracyof98%forthe
microPHAZIRTMleafmodel(Model1)and97.4%forthemicroPHAZIRTMberrymodel(Model2).
Models1and2presentedvalidationaccuracies(94%and93%,respectively)closetothoseofthe
trainingstage(100%bothmodels).Furthermore,performancevaluesfortraining(Models1and2:
MSE<0.01)werelowerthantheotherstagesandvalidation(Model1:MSE=0.02;Model2:MSE
=0.03)andtesting(Model1:MSE=0.02;Model2:MSE=0.04)weresimilar;thisindicatesthatthere
werenosignsofoverfittingforbothModel1andModel2.
Figure7depictsthereceiveroperatingcharacteristic(ROC)curvesforthetwoANNmodels
developed.Allmodelsshowedhightruepositiverates(sensitivity)andlowfalsepositiverates
(specificity)forclassifyingthespectralreadingsaccordingtotheexperimentaltreatment,whichcan
alsobeobservedinthelastcolumnofeachconfusionmatrix.ForModel2,theHStreatmenthadthe
highestsensitivity(100%),followedbytheCMandHSMtreatments(99.1%each)andLStreatment
(96.3%).TheCtreatmenthadthelowestsensitivityof92.6%forthismodel.ForModel1,theC
treatmenthadthehighestsensitivity(99.1%),followedbytheLStreatment(98.8%),HStreatment
(97.8%),andCMtreatment(97.5%),whiletheHSMhadthelowestsensitivityof96.9%.
Table4.Statisticalresultsfortheartificialneuralnetworkspatternrecognitionmodels.Model1:
microPHAZIRTMforleaves,andModel2:microPHAZIRTMforberries.Performanceisbasedon
meanssquarederror(MSE).
StageSamples(n)Accuracy%Error%Performance(MSE)
Model1
Training113110000.00
Validation 24394.25.80.02
Testing24392.67.40.02
Overall161798.02‐
Model2
Training37810000.00
Validation 8192.67.40.03
Testing8190.19.90.04
Overall54097.42.6‐
Sensors2020,20,xFORPEERREVIEW16of24
(a)
(b)
Figure7.Receiveroperatingcharacteristic(ROC)curvesforthetwomodelsdeveloped(a)the
microPHAZIR
TM
leafmodel,(b)themicroPHAZIR
TM
berrymodel.Coloredlinesrepresentthe
differentsmokingtreatments.Abbreviations:C=controlwithoutmisting;CM=controlwithmisting;
HS=highdensitysmokewithoutmisting;HSM=highdensitysmokewithmisting;andLS=low
densitysmoke.
Sensors2020,20,xFORPEERREVIEW17of24
4.Discussion
4.1.PhysiologicalMeasurements
Leafgasexchangeparametersweremeasuredthedayaftersmoking.Thethreesmoke
treatmentsshowedsignificantreductionsings,inparticular,thehighdensitysmokewithoutmisting
(HS)treatment,whichshowedthelowestaveragereadingforgs(Table2).Stomatalclosureisoneof
thefirstresponsestosmokeexposureundertakenbyplants[6,26],andastudybyRisticand
colleagues[26]foundthatthetimerequiredforgstorecoverfollowingonehourofsmokeexposure
forCabernetSauvignongrapevineswasapproximately6–10days.ApreviousstudybyBelletal.[6]
foundthatgsofpottedCabernetSauvignongrapevineshadreturnedto60%ofpresmokeexposure
ratefollowingfifteenminexposuretosmokeusingTasmanianbluegum(EucalyptusglobulusL.)
leavesasfuel,whilerateshadreturnedto80%ofpresmokevaluesfollowingexposuretosmoke
derivedfromCoastLiveOak(QuercusagrifoliaNée)leaves.Thisindicatesthatinadditiontothetype
offuelused,theintensityofsmokeexposuremayalsoaffecttheextentofstomatalclosureand,hence,
reductionings.Itis,therefore,notsurprisingthattheHStreatmenthadthelowestgs.However,itis
interestingthatthelowsmoketreatment(LS)hadlowergsthanthehighsmokewithmisting
treatment(HSM),whichindicatesthatmistingmayhavereducedtheeffectofsmokeexposureongs.
Duringabushfire,thetypeoffuelburntwillvarydependingontheregionandthetypeofplant
speciesnativetothearea,aswellastheamountofsmokeexposureduetolandtopographyandwind
vectors;therefore,theeffectongsmayvary[17,18,23,66].Whilemistingonlypartiallypreventedthe
uptakeofvolatilephenolsandglycoconjugatesingrapes[60],itdidappeartohaveaphysiological
effect.Itisevidentthatmistingreducedtheeffectofsmokeexposureongs.Smokecontainsacomplex
mixtureofgasessuchassulfurdioxide(SO2),O3,andnitrogendioxide(NO2),aswellasdustparticles
thathavebeenshowntoinhibitphotosynthesisandaffectstomatalopening[6,26,29].Stomataarethe
primarypointofentryforthesegasesanddustparticles[6];therefore,mistingmayhelppreventthe
uptakeofdustandotherparticlesbytrappingtheminwaterthathascondensedontheleafsurface,
preventingtheirentranceintothestomata.Thepresentwatermayalsoactasasolventforgasessuch
asSO2andNO2,therebyincorporatingthemintoasolutionthatthenmaydripofftheleafsurface.In
additiontothis,smokeexposuremaytriggerstomatalclosurebyproducinghighvaporpressure
deficits[26,29].Thepresenceofmistingmayhelpreducetheleaftoairvaporpressuredifference
producedbysmokeexposure,therebyreducingtheimpactongs.Mistingalsoappearedtoreduce
theeffectofsmokeexposureontranspirationrate(E)astherewerenodifferencesbetweenthetwo
controltreatmentsandtheLSandHSMtreatments.OnlytheHStreatmenthadsignificantlyreduced
E.Meanratesofphotosynthesis(A)followedsimilarpatternstogs,withtheHStreatmenthavingthe
lowestvalue,followedbytheLStreatmentandthentheHSMtreatment,whilethecontrolwithout
misting(C)hadthehighestrateofA.ThisindicatesthatwhilemistingmayhavereducedAinthe
controltreatments,itmayalsohelpreducetheeffectsofsmokeexposureonA.
4.2.NearInfraredSpectroscopyPatternsandPrincipalComponentAnalysis
FromthePCAbiplots(Figures5and6)andspectra(Figures3and4)generatedinthecurrent
study,itisevidentthatsmokeexposurealterstheNIRspectralsignalsofgrapevineleavesand
berries,andthismayproveusefulforthedetectionofgrapevinesmokecontamination.Forthe
microPHAZIRTMRXAnalyzerleafspectra,highloadings(Figure5b.)wereobservedforthe
wavelengthregionsbetween1604–1621,1621–1647,and1613–1647nm,allofwhichcorrespondtoC
Hstretchingofsugarsandaromaticcompounds[67–70].ForthemicroPHAZIRTMRXAnalyzerberry
spectra,highloadings(Figure6b.)wereobservedforthewavelengthregionsbetween1604–1622,
1630–1647,and2374–2389nm,whichcorrespondtoCHstretchingofsugars,suchasglucose,aswell
asaromatichydrocarbons,whichmaybeduetothepresenceofsmokederivedvolatilephenols,such
asguaiacols,cresols,andsyringols,andtheirglycoconjugates[67,68,71,72].
Sensors2020,20,xFORPEERREVIEW18of24
4.3.SpectralIndices
4.3.1.Leaf
Thenormalizeddifferencevegetationindex(NDVI)givesanindicationofplantvigorandfruit
ripeningresultingfromrelativechangesinchlorophyllcontent.Itisbasedonthevariationbetween
themaximumabsorptionofredbychlorophyllpigmentsandthemaximumreflectanceinthe
infraredcausedbyleafcellularstructure[56,57,73–75].Similarly,relativechangesinanthocyanin
contentareexpressedasthenormalizedanthocyaninindex(NAI).BoththeNDVIandNAIare
expressedasanormalizedvaluebetween−1(lackofgreenorredness)to+1(greenorred)[56,57].
Notsurprisingly,HSleaveshadthelowestNDVIandNAIvalues.Previousstudiesinvestigatingthe
effectsofpollutiononleafpigmentsfoundadecreaseinphotosyntheticpigmentsfollowingexposure
topollutants,includingsulfurdioxide(SO2),carbondioxide(CO2),nitrogendioxide(NO2),andozone
(O3)[59,76,77].Thesestudiesareoftenusedascomparisonsforinvestigatingtheeffectsofsmoke
exposureonleavesascompoundsinairpollutioncanalsofoundinsmoke[6,22].Therewereno
differencesinNDVIandNAIvaluesbetweentheLS,HSM,andcontroltreatments(CandCM),
indicatingthatmistingmayreducetheeffectsofsmokeexposureonleafpigments,andlowlevelsof
smokeexposureforonehourmayalsohavenoeffect.Longerperiodsofsmokeexposure(daysor
weeks,asisoftenthecasewithwildfires)mayberequiredtocauseanoticeablechangeinleaf
pigments.
Theplantsenescencereflectanceindex(PSRI)givesanindicationofthestageofleafsenescence
andfruitripeningthroughassessingchangesincarotenoidaccumulationandtheirproportionto
chlorophyll.Valuesrangefrom−1to+1,withhighervaluesindicatingincreasedstressandcarotenoid
accumulation[55,63–65,78].ThePSRIwashighestfortheHStreatment,indicatingheightenedstress
andleafsenescence.ThisalsocorrespondswiththehighCRI500valueforthissmoketreatment,
indicatingincreasedcarotenoidaccumulation.
4.3.2.Berries
ResearchbyNoesthedenetal.[5]foundthatsmokeexposureinducedchangesin
phenylpropanoidmetabolitesinPinotNoirberriesandwine,someofwhichareassociatedwiththe
colorandmouthfeelofthewine.BerriesexposedtoHSandLStreatmentshadthehighestmeanNAI
values,indicatingthatsmokeexposuremayincreaseanthocyanincontent,possiblyduetoanincrease
inphenolicaccumulationasastressresponseinducedbyexposuretoozonepresentinsmoke
[5,79,80].TheHSMtreatmenthadalowNAIvalue,indicatingthatmistingmayreduceanthocyanin
concentrationsthroughincreasedirrigation.Castellarinetal.[81]foundthatearly(beforevéraison)
andlate(aftertheonsetofripening)season,waterdeficitsincreasedanthocyaninaccumulation
duringripening.Theapplicationofincanopymistingmayreducewaterstressand,therefore,reduce
anthocyaninaccumulation.
InterestinglytheHSMfollowedbytheHStreatmentshadthelowestPSRIvalues.Ascarotenoid
concentrationsingrapesgenerallydecreaseduringvéraison,thismayhaveresultedinlowerPSRI
values.Therefore,thePSRImaynotbesuitableforassessingthedegreeofripeningingrapeberries.
4.4.ANNModeling
BothANNmodelsclassifiedleafandberryreadingsasafunctionofsmokeexposurewithhigh
accuracy.ThemicroPHAZIRTMleafmodel(model1)hadthehighestpositiveclassification,with98%
accuracy(Table4).TheNIRregionselectedforuseinModel1wasbetween1600–1800nminorder
tominimizeanypossibleinterferenceduetotheabsorptionspectraofwaterintheregionof
approximately1930nm[69].Furthermore,theregionbetween1680–1690nmisassociatedwith
aromaticCHstretching[67];assuch,anypatternsobservedbytheANNwouldmostlikelybedue
tothepresenceofsmokederivedvolatilephenols.ResearchbyKennison[22]foundapositive
correlationbetweenlevelsofsmokederivedcompoundsfoundinleavesandlevelsinwine;this
ANNmodeldevelopedmay,therefore,offerarapid,infieldmethodforassessinggrapevinesmoke
Sensors2020,20,xFORPEERREVIEW19of24
contamination.ItalsodemonstratesgreatpromiseforfurtherresearchintotheuseofNIR
spectroscopycoupledwithunmannedaerialvehicles(UAVs)withGlobalPositioningSystem(GPS)
trackers,whichcouldflyovervineyardstoscangrapevinecanopiesandprovidemapsofsmoke
contaminatedregions.
ThemicroPHAZIRTMberrymodel(model2)alsohadahighoverallaccuracyinclassifyinggrape
berriesaccordingtosmoketreatment(97.4%).ForModel2,theentirewavelengthrangebetween
1600–2396nmwasused.ThisincludestheCHstretchingofaromaticcompoundsat1680nm,OH
stretchingat1930nmassociatedwithglucose,cellulose,andwater,andC=Osecondovertone
associatedwithcarboxylicacidsandwaterbetween1900–1910nm[67,69].AsNIRmeasurements
wereconductedinfieldonwholeberries,thisoffersanondestructivetoolforassessinggrapevine
smokecontamination.Wholegrapesmaybeusedforassessmentassmokecompoundshavebeen
foundtooccurprimarilyingrapeskins[3,25].Furthermore,theLightingPassportTMsmarthandheld
spectrometermaybeofinteresttogrowersduetoitsaffordabilitycomparedtootherspectrometers.
Itisalsoverysmallandlightweight,makingiteasytoundertakemeasurementsinfield,anditcan
beconnectedtosmartphonesviaBluetooth,wheredatacanbestoredandretrievedforlateranalysis
[82].
ThetwoANNmodelsmoreaccuratelydifferentiatedthespectralreadingsrelativetoPCA.This
maybebecauseANNsarebettersuitedtohandlecomplex,nonlineardata,andmorereadilyfind
patternsorrelationshipsbetweendatathanotherformsofanalysis[53,83–85].ResearchbyJaniket
al.[53]foundthatthecombinationofANNswithpartialleastsquares(PLS)orPCAovercomesissues
ofnonlinearityaswellasincreasingtheaccuracyofregressionmodelsinpredictingtotal
anthocyaninconcentrationsinredgrapehomogenates.ThismayalsoexplainwhyModel2wasable
toaccuratelydifferentiatetheberryspectralreadingsfromC,CM,andLStreatments,despiteanalysis
ofvarianceindicatingtherewerenostatisticallysignificantdifferences.
Assmokeexposurealteredthechemicalfingerprintingofgrapevineleavesandberries,theANN
modelswereabletodetectchangesinthespectralpatternsandthenclassifythereadingsasafunction
ofexperimentaltreatments.Thismayoffergrapegrowersarapidmethodofassessingthelevelof
smokecontaminationingrapeberriesandleaves,withahighlevelofaccuracyandprecision.This
mayassistgrowersindecidingwhichberrysamplestosendforfurtherchemicalanalysistoquantify
thelevelsofsmokecompoundsingrapesandpredictthelevelofsmoketaintinthefinalwine,or
theymaydecidetoavoidharvestingheavilycontaminatedgrapesforwinemaking.Furthermore,as
thismethodisnondestructive,repeatedmeasurementsarepossible.Byknowingthelevelofsmoke
contamination,growerscanmakeinformeddecisions.
WhiletheANNmodelsdevelopedwereabletoclassifyCabernetSauvignonleafandberry
spectraaccurately,furtherresearchisrequiredtoassesswhetherthesemodelscanbeusedforother
grapevarieties,asdifferencesinberrycompositionandleafphysiologymayaffecttheaccuracyof
classification[6,34].PreviousresearchevaluatedMIRspectroscopyfortheclassificationofsmoke
taintedwinesfoundcompositionaldifferencesduetograpevarietyprevailedoverdifferences
resultingfromlowlevelsofsmokeexposure[34].Furthermore,thephysiologicalresponsesof
differentgrapevarietiestosmokewerefoundtovary,bothinmagnitudeandinrecoverytime[6,26].
Thus,furthertestingofthesemodelsusingberryandleafspectrafromdifferentgrapevinevarieties
isrequired.
5.Conclusions
ResultsfromthisstudyindicatethatsmokeexposurealterstheNIRspectraofCabernet
Sauvignongrapevineleavesandberries.Asaresult,accurateclassificationmodelscanbedeveloped
usingANNmodeling.Artificialneuralnetworksarebetteratclassifyingnonlinearorcomplexdata
thantraditionaltechniques,suchasprincipalcomponentanalysis.Furthermore,theuseofUVVis
spectroscopymayofferinsightsintothephysiologicalperformanceofleavesandthequalityand
degreeofripeningofgrapes.Thesetechniquesmayassistgrapegrowersinidentifyinggrapevines
thathavebeencontaminatedbysmoke,therebyinformingdecisionmakingtoavoidharvestingand
processingheavilycontaminatedgrapesand/ortheneedformitigationtechniquestomanagetherisk
Sensors2020,20,xFORPEERREVIEW20of24
ofsmoketaintinresultingwine.FurthertestingoftheANNmodelsdevelopedinthecurrentstudy
isrequiredtoassesstheiraccuracyinclassifyinggrapevineleafandberryspectrafromothergrape
varieties.
SupplementaryMaterials:Thefollowingareavailableonlineatwww.mdpi.com/14248220/20/18/5099/s1,Table
S1:Concentrationsofvolatilephenolsingrapejuice(μg/L)andtheirglycoconjugatesingrapehomogenate
(μg/kg)onehouraftersmoketreatments.
AuthorContributions:Conceptualization,V.S.,andS.F.;datacuration,V.S.,C.G.V.,andS.F.;formalanalysis,
V.S.;fundingacquisition,S.F.;investigation,V.S.,andC.G.V;methodology,V.S.,C.G.V.,C.S,K.L.W.,andS.F.;
projectadministration,K.L.W.,andS.F.;resources,K.L.W.,R.D.B.,andS.F.;software,C.G.V.,andS.F.;
supervision,D.D.T.,A.P.,andS.F.;validation,C.G.V.,K.L.W.,andS.F.;visualization,V.S.,C.G.V.,D.D.T.,and
S.F.;writing—originaldraft,V.S.;writing—reviewandediting,V.S.,C.G.V.,C.S.,K.L.W.,D.D.T.,A.P.,R.D.B.,
andS.F.Allauthorshavereadandagreedtothepublishedversionofthemanuscript.
Acknowledgments:ThisresearchwassupportedthroughtheAustralianGovernmentResearchTraining
ProgramScholarship,aswellastheDigitalViticultureprogramfundedbytheUniversityofMelbourne’s
NetworkedSocietyInstitute,Australia.C.S.wassupportedbytheAustralianResearchCouncilTrainingCentre
forInnovativeWineProduction(www.arcwinecentre.org.au),whichisfundedaspartoftheARC’sIndustrial
TransformationResearchProgram(ProjectNo.ICI70100008),withsupportfromWineAustraliaandindustry
partners.TheauthorsgreatlyacknowledgetheDigitalAgriculture,Food,andWineGroup.
Funding:Thisresearchreceivednoexternalfunding.
ConflictsofInterest:Theauthorsdeclarenoconflictofinterest.Thefundershadnoroleinthedesignofthe
study;inthecollection,analyses,orinterpretationofdata;inthewritingofthemanuscript,orinthedecisionto
publishtheresults.
References
1. Cain,N.;Hancock,F.;Rogers,P.;Downey,M.Theeffectofgrapevarietyandsmokingdurationonthe
accumulationofsmoketaintcompoundsinwine.WineVitic.J.2013,28,48–49.
2. CSIRO.AustralianGovernmentBureauofMeteorology.StateClim.2018,2018,5.
3. Dungey,K.A.;Hayasaka,Y.;Wilkinson,K.L.Quantitativeanalysisofglycoconjugateprecursorsofguaiacol
insmokeaffectedgrapesusingliquidchromatography–tandemmassspectrometrybasedstableisotope
dilutionanalysis.FoodChem.2011,126,801–806.
4. Kennison,K.;Gibberd,M.;Pollnitz,A.;Wilkinson,K.Smokederivedtaintinwine:Thereleaseofsmoke
derivedvolatilephenolsduringfermentationofmerlotjuicefollowinggrapevineexposuretosmoke.J.
Agric.FoodChem.2008,56,7379–7383.
5. Noestheden,M.;Noyovitz,B.;RiordanShort,S.;Dennis,E.G.;Zandberg,W.F.Smokefromsimulated
forestfirealterssecondarymetabolitesinVitisviniferaL.Berriesandwine.Planta2018,248,1537–1550.
6. Bell,T.;Stephens,S.;Moritz,M.Shorttermphysiologicaleffectsofsmokeongrapevineleaves.Int.J.
WildlandFire2013,22,933–946.
7. DeVries,C.;Mokwena,L.;Buica,A.;McKay,M.DeterminationofvolatilephenolinCabernetsauvignon
wines,madefromsmokeaffectedgrapes,byusinghsspmeGCMS.S.Afr.J.Enol.Vitic.2016,37,15–21.
8. Noestheden,M.;Dennis,E.G.;RomeroMontalvo,E.;DiLabio,G.A.;Zandberg,W.F.Detailed
characterizationofglycosylatedsensoryactivevolatilephenolsinsmokeexposedgrapesandwine.Food
Chem.2018,259,147–156.
9. Parker,M.;Osidacz,P.;Baldock,G.A.;Hayasaka,Y.;Black,C.A.;Pardon,K.H.;Jeffery,D.W.;Geue,J.P.;
Herderich,M.J.;Francis,I.L.Contributionofseveralvolatilephenolsandtheirglycoconjugatestosmoke
relatedsensorypropertiesofredwine.J.Agric.FoodChem.2012,60,2629–2637.
10. Kennison,K.;Wilkinson,K.;Pollnitz,A.;Williams,H.;Gibberd,M.Effectofsmokeapplicationtofield
grownMerlotgrapevinesatkeyphenologicalgrowthstagesonwinesensoryandchemicalproperties.
Aust.J.GrapeWineRes.2011,17,5–12.
11. Ristic,R.;vanderHulst,L.;Capone,D.;Wilkinson,K.Impactofbottleagingonsmoketaintedwinesfrom
differentgrapecultivars.J.Agric.FoodChem.2017,65,4146–4152.
12. Härtl,K.;Schwab,W.Smoketaintinwinehowsmokederivedvolatilesaccumulateingrapevines.Wines
Vines2018,99,62–64.
Sensors2020,20,xFORPEERREVIEW21of24
13. Hayasaka,Y.;Baldock,G.;Pardon,K.;Jeffery,D.;Herderich,M.Investigationintotheformationofguaiacol
conjugatesinberriesandleavesofgrapevineVitisviniferaL.Cv.Cabernetsauvignonusingstableisotope
tracerscombinedwithhplcmsandms/msanalysis.J.Agric.FoodChem.2010,58,2076–2081.
14. Hoj,P.;Pretorius,I.;Blair,R.Investigationsconductedintothenatureandameliorationoftaintsingrapes
andwine,causedbysmokeresultingfrombushfires.Aust.WineRes.Inst.Annu.Rep.2003,37–39.
15. Kelly,D.;Zerihun,A.;Singh,D.;VitzthumvonEckstaedt,C.;Gibberd,M.;Grice,K.;Downey,M.Exposure
ofgrapestosmokeofvegetationwithvaryinglignincompositionandaccretionofligninderivedputative
smoketaintcompoundsinwine.FoodChem.2012,135,787–798.
16. Singh,D.;Chong,H.;Pitt,K.;Cleary,M.;Dokoozlian,N.;Downey,M.Guaiacoland4methylguaiacol
accumulateinwinesmadefromsmokeaffectedfruitbecauseofhydrolysisoftheirconjugates.Aust.J.
GrapeWineRes.2011,17,S13–S21.
17. Noestheden,M.;Thiessen,K.;Dennis,E.G.;Tiet,B.;Zandberg,W.F.Quantitatingorganolepticvolatile
phenolsinsmokeexposedVitisviniferaberries.J.Agric.FoodChem.2017,65,8418–8425.
18. Krstic,M.;Johnson,D.;Herderich,M.Reviewofsmoketaintinwine:Smokederivedvolatilephenolsand
theirglycosidicmetabolitesingrapesandvinesasbiomarkersforsmokeexposureandtheirroleinthe
sensoryperceptionofsmoketaint.Aust.J.GrapeWineRes.2015,21,537–553.
19. Kennison,K.;Wilkinson,K.;Williams,H.;Smith,J.;Gibberd,M.Smokederivedtaintinwine:Effectof
postharvestsmokeexposureofgrapesonthechemicalcompositionandsensorycharacteristicsofwine.J.
Agric.FoodChem.2007,55,10897–10901.
20. Claughton,C.;Jeffery,C.;Pritchard,M.;Hough,C.;Wheaton,C.WineIndustry’s‘BlackSummer’asCost
ofSmokeTaint,BurntVineyards,andLostSalesAddup.ABCNews,28February2020.
21. Dutta,R.;Das,A.;Aryal,J.BigdataintegrationshowsAustralianbushfirefrequencyisincreasing
significantly.R.Soc.OpenSci.2016,3,150241.
22. Kennison,K.BushfireGeneratedSmokeTaintinGrapesandWine.FinalReporttoGrapeandWineResearchand
DevelopmentCorporation;RD05/02–3;DepartmentofAgricultureandFoodWesternAustralia:Kalgoorlie,
Australia,2009.
23. Simos,C.Theimplicationsofsmoketaintandmanagementpractices.Aust.Vitic.Jan./Feb.2008,12,77–80.
24. Fuentes,S.;Tongson,E.Advancesinsmokecontaminationdetectionsystemsfor