You will need the Flash plugin to view this publication

Download the Flash plugin for free here. Just refresh this page after installing Flash.
Also make sure that javascript is enabled in your browser.

Alternatively, you can download this publication in PDF-format - Click here.

Page 1
Page 2
Page 3
Page 4
Page 5
Page 6
Page 7
Page 8
Page 9
Page 10
Page 11
Page 12
Page 13
Page 14
Page 15
Page 16
Page 17
Page 18
Page 19
Page 20
Page 21
Page 22
Page 23
Page 24
Page 25
Page 26
Page 27
Page 28
Page 29
Page 30
Page 31
Page 32
Page 33
Page 34
Page 35
Page 36
Page 37
Page 38
Page 39
Page 40
Page 41
Page 42
Page 43
Page 44
Page 45
Page 46
Page 47
Page 48
Page 49
Page 50
Page 51
Page 52
Page 53
Page 54
Page 55
Page 56
Page 57
Page 58
Page 59
Page 60
Page 61
Page 62
Page 63
Page 64
Page 65
Page 66
Page 67
Page 68
Page 69
Page 70
Page 71
Page 72
Page 73
Page 74
Page 75
Page 76
Page 77
Page 78
Page 79
Page 80
Page 81
Page 82
Page 83
Page 84
Page 85
Page 86
Page 87
Page 88
Page 89
Page 90
Page 91
Page 92
Page 93
Page 94
volume21number12013ISSN0967-0335INTHISISSUEpowderblendingprocessanalysisbackscatteringandtransmissionoflightbymilkdesignedorthogonalsamplespikingbasedcalibrations JOURNALOFNEARINFRAREDSPECTROSCOPYEditor-in-chiefGraemeD.BattenSeaSpecPtyLtdPOBox487WoolgoolgaNSW2456Australia.PhoneFax61-266-562288.E-mailthebattensbigpond.comAdjunctProfessoratCharlesSturtUniversityHonoraryProfessoratTheUniversityofSydneyEditorsFranklinE.BartonIILightLightSolutionsLLCPOBox81486AthensGA30608-1486USA.E-mailbartonlightlightsolutions.comGerardDowneyAshtownFoodResearchCentreAshtownDublin15Ireland.E-mailgerard.downeyteagasc.ieTomFearnDepartmentofStatisticalScienceUniversityCollegeLondonGowerStreetLondonWC1E6BTUK.E-mailt.fearn.ucl.ac.ukChristianHuckInstituteofAnalyticalChemistryandRadiochemistryCCB-CenterforChemistryandBiomedicineLeopold-FranzensUniversityInnrain80-826020InnsbruckAustria.E-mailchristian.w.huckuibk.ac.atSumioKawanoFacultyofAgricultureKagoshimaUniversity1-21-24KorimotoKagoshimacity890-0065Japan.E-mailkawanoagri.kagoshima-u.ac.jpRogerMederTropicalProductionForestryCSIRO306CarmodyRoadStLuciaQueensland4067AustraliaKarlH.NorrisConsultant11204MontgomeryRoadBeltsvilleMD20705USA.E-mailknnirsgmail.comBradSwarbrickCAMOSoftwareASUSA.E-mailbswarbrickcamo.comFoundingeditorAnthonyM.C.DaviesNorwichUK.E-mailtdnnirc.co.ukEditorialadvisoryboardP.DardenneLibramontBelgiumE.W.CiurczakLaurelMDUSAM.FerrariLAquilaItalyP.FlinnVictoriaAustraliaA.Garrido-VaroCrdobaSpainD.W.HopkinsBattleCreekMIUSAM.IwamotoIbarakiJapanM.ManleyStellenboschRSAH.MarkSuffernNYUSAB.G.OsborneNorthRydeAustraliaY.OzakiSandaJapanH.W.SieslerEssenGermanyP.C.WilliamsNanaimoCanadaSubmissionofarticlesAuthorsintendingtosubmitanarticleforpublicationshouldconsulttheGuidelinesforAuthorsavailablefromthePublishersorEditors.Forthefastestdecisionmanuscriptsshouldbesubmittedviaouronlinesystemathttpwww.impublications.comauthors.TheymayalsobesenttotheEditor-in-ChiefortoanEditorconvenienttotheauthor. Thisdisksymbolontherstpageofanarticleindicatestheauthorsswillingnesstomakeavailablethedata.Ifyouwantityoushouldcontacttheauthorswitharequestforthedataandtheundertakingthatitwillbeusedforresearchpurposesandnotusedinanypublicationwithouttheiragreement.Anasteriskagainstanauthorsnameontherstpageofanarticleindicatestheauthortowhomcorrespondenceshouldbeaddressed.JOURNALOFNEARINFRAREDSPECTROSCOPYContentsEditorialvGraemeD.BattenTabletcharacteristicspredictionbypowderblendingprocessanalysisbasedonnearinfraredspectroscopy1YusukeHattoriYoshiyukiTajiriandMakotoOtsukaDesignedorthogonalsamplespikingbasedcalibrationsforquantitativeliquidphasemeasurementswithnearinfraredspectroscopyinananaerobicdigestionprocess11JohnDahlbackaTomLillhongaandMarcoDringQuantitativemeasurementsofanaerobicdigestionprocessparametersusingnearinfraredspectroscopyandlocalcalibrationmodels23JohnDahlbackaandTomLillhongaShort-wavenearinfraredspectrometryofbackscatteringandtransmissionoflightbymilkformulti-componentanalysis35AndreyV.KalininViktorN.KrasheninnikovandVladimirM.KrivtsunRapidandquantitativedetectionmethodforacteosideduringchromatographicpuricationofadhesiverehmannialeafextractusingnearinfraredspectroscopyandchemometrics43YeJinXuesongLiuLianjunLuanGuangmingQingYingZhongandYongjiangWuAcomprehensivenearinfraredspectroscopicstudyofthelimitsofquantitativeanalysisofsulfathiazolepolymorphism55PlMacFhionnghaileYunHuPatrickMcArdleandAndreaErxlebenComparingpredictiveabilitiesofthreevisible-nearinfraredspectrophotometersforsoilorganiccarbonandclaydetermination67MariaKnadelBoStenbergFanDengAntonThomsenandMogensHumlekrogGreve ISSN0967-0335IMPublicationsLLP6CharltonMillCharltonChichesterWestSussexPO180HYUK.Tel44-01243-811334Fax44-01243-811711Webwww.impublications.comAimsandscopeTheJournalofNearInfraredSpectroscopyaimstopublishoriginalresearchpapersshortcommunicationsreviewarticlesandlettersconcernedwithnearinfraredspectroscopyandtechnologyitsapplicationnewinstrumentationandtheuseofchemometricanddatahandlingtechniqueswithinNIR.InthiswayitintendstoserveasaforumfortheexchangeofinformationonallaspectsofNIRspectroscopyandtechnology.TheJournalofNearInfraredSpectroscopywillacceptcontributionsfromallareasofapplicationwherenearinfraredspectroscopyisinusesuchasAgricultureChemicalIndustryFoodLifeSciencesProcessControlPharmaceuticalsTextilesandPolymers.TopicsforpapersmayincludeChemometricsCalibrationsDiffuseReectionNIRImagingOn-LineUseFibreOpticsSamplingSpectroscopyInstrumentationandRemoteSensing.IMPublicationsLLP2013Allrightsreserved.ApartfromanyfairdealingforthepurposesofresearchorprivatestudyorcriticismorreviewaspermittedundertheUKsCopyrightDesignsandPatentsAct1988thispublicationmaybereproducedstoredortransmittedinanyformorbyanymeansonlywiththepriorpermissioninwritingofthepublishersorinthecaseofreprographicreproductioninaccordancewiththetermsoflicencesissuedbytheCopyrightLicensingAgency.Enquiriesconcerningreproductionoutsidethosetermsshouldbesenttothepublishersattheaddressabove.SubscriptiondetailsJournalofNearInfraredSpectroscopyispublishedbimonthly.Thesubscriptionratesfortheprintandonlineeditionare344.00deliveryaddressesinUK431.00deliveryincontinentalEuropeorUS546.00deliveryaddressesoutsideEuropefor2013.SubscriptionscanonlybeacceptedforafullcalendaryearandshouldbeorderedfromthepublishersIMPublicationsLLP6CharltonMillCharltonChichesterWestSussexPO180HYUKTel4401243811334Fax4401243811711e-mailsubsimpublications.co.ukorfromyoursubscriptionagent.ThesubscriptionratesincludeairmaildeliverytoallcountriesexcepttheUKwherenormalinlandpostageisused.AreducedratesubscriptionpackageisavailabletosubscriberstakingbothJournalofNearInfraredSpectroscopyandNIRnewsthisonlyapplieswhenbothpublicationsareorderedandpaidforwithasingletransaction.AnelectronicversionofJournalofNearInfraredSpectroscopyisalsoavailableonCD-ROMandontheworldwideweb.Furtherdetailsareavailablefromthepublishersattheaddressabove.AbstractingJournalofNearInfraredSpectroscopyisabstractedinChemicalAbstractsScienceCitationIndexExpandedSCIEResearchAlertChemistryCitationIndexCCICurrentContentsPhysicalChemicalEarthSciencesCCPCESCurrentContentsAgricultureBiologyEnvironmentalSciencesCCABES.AdvertisingAdvertisingisacceptedinJournalofNearInfraredSpectroscopyfurtherdetailscanbeobtainedfromtheAdvertisementManagerIMPublicationsLLP6CharltonMillCharltonChichesterWestSussexPO180HYUKTel4401243-811334Fax4401243-811711e-mailianimpublications.co.uk.PrintedintheUKbyLatimerTrendCompanyLtdPlymouth. JOURNALOFNEARINFRAREDSPECTROSCOPYvISSN0967-0335IMPublicationsLLP2013doi10.1255jnirs.1039AllrightsreservedThe20thanniversaryvolumeofJNIRSJournalofNearInfraredSpectroscopycontained59papersincluding14reviewpapersissues1and5werespecialissuesdevotedtomedicalandimagingmattersrespectively.At706pagesofpeer-reviewedpapersVolume20wasourlargestyetIammostgratefultotheGuestEditorsMarcoFerrariKarlNorrisMikeSowaMarenaManleyandGeraldDowneyfortheirexcellentworkinencouragingauthorstocontributepaperstothesespecialissuesandthenoverseeingtheirreviewandrevision.Ialsowishtoacknowledgethededicatedeffortsofallwhoreviewpapers.TheircommentsandsuggestionshelpustomaintainthehighstandardwhichreadersofJNIRSexpect.Priortopublicationthepapersarecarefullyeditedandtypeset.AtthisstageMsGillStockfordensuresthattexttablesandguresarepresentedtoreadersatamostprofessionalstandard.ThepaperspublishedinVolume20containanarrayofnovelandinterestingndingswhichenhanceourunderstandingofandcondenceintheabilityofnearinfraredspectroscopytoanalysematerialsrapidlyaccuratelyandatalowercostthanwaspreviouslypossible.ThereisacontinuingowofadvancesinourknowledgeofhowlightinteractswithmatterincludinghowbesttointerpretthedatainspectraoftheNIRwavelengths.TheseadvancesarebeingmadeatboththebasiclevelofNIRspectroscopyandwherethisknowledgeisappliedacrosstheworld.ManyapplicationsofNIRspectroscopyhaveevolvedtomeettheneedsofindustriesforexampletoraisecropyieldsandsegregatecerealsaccordingtoqualityproteintoprotectpeoplefromtoxinsforexampletodetectdangerouspathogenssuchasE.coliandFusariumonfoodstodetectcontaminatedfeedstuffsbydetectingmeatandbonemealinanimalfoodsandtoenableveryhighqualitycontrolsuchascheckingthelevelsanddistributionofactiveingredientsinpharmaceuticaltablets.ItwasclearlyevidentinthemedicalspecialissuethatNIRspectroscopyisabletomonitortheoxygenstatusofbloodandnon-invasivelyexplainarangeofhumanhealthproblems.TheuseofNIRspectroscopytodetectcounterfeitandincorrectlylabelledgoodsisalsoavaluableuse.ThesearejustafewexamplesofhowNIRspectroscopyismakingadifferenceandisagreatbenettomankind.Ibeginmy10thyearasEditor-in-ChiefofJNIRSwithhighexpectationsthat2013willrevealinterestingandvaluableknowledgewhichenhancesourabilitytoanalysesampleswhethertheybeanimalvegetableormineralwithmorecondenceandtothebenetofmorepeopleandtheenviron-ment.ThisrstissueofJNIRSfor2013includesapaperwhichreportsbasicNIRscience.A.V.KalininV.N.KrasheninnikovandV.M.Krivtsun1challengeourunderstandingofhowbesttodeterminethecompositionofacomplexsubstancelikemilk.Theneedforhumanstoturnwasteproductsintore-usableproductshasstimulatedNIRresearchintheareaofanaerobicdigestionoforganicmaterials.ThepapersbyJohnDahlbackaandTomLillhonga23areexamplesofthecontributionsofNIRscientiststobothmorereliabledatafromNIRanalysesandthemonitoringofprocessesinrealsituations.ItisacredittoNIRscientiststhattheycontinuetoseekaclearerunderstandingofhowlightinteractswithmatterandhowthisknowledgecanbeapplied.JNIRSisproudtopublishthisnewknowledge.Ihaveaneditorialteamwhichisdedi-catedtoexcellenceandIampleasedthatProfessorChristianHuckDrRogerMederandDrBradSwarbrickhaverecentlyjoinedJNIRSasspecialisteditors.ThereisaworldwidetrendtowardsOpenAccesspublication.JNIRShasforsomeyearsofferedauthorstheoptiontopublishpapersusingtheOpenAccessmodel.ThesepapersaretreatedinexactlythesamewayasotherJNIRSpapersthroughthecopyeditingandtypesettingprocessbutarefreelyavailabletoallandauthorshavepermissiontoplacethemwheretheywishsuchasontheirpersonalwebpagesandininstitutionalorotherrepositories.IneithercasetheEditorialGraemeD.BattenEditor-in-ChiefSeaSpecPtyLtdPOBox487WoolgoolgaNSW2456Australia.E-mailthebattensbigpond.comG.D.BattenJ.NearInfraredSpectrosc.21vvi2013 viEditorialpapersaresubjectedtoandmustmeetthehighstandardofsciencethatreadersofJNIRShavecometoexpect.IndeedweaskauthorsnottorequesttheOpenAccessoptionuntiltheirpaperhasbeenaccepted.Formoreinformationseewww.impublications.comopen-access.References1.A.V.KalininV.N.KrasheninnikovandV.M.KrivtsunShortwavenearinfraredspectrometryofbackscatteringandtransmissionoflightbymilkformulti-componentanalysisJ.NearInfraredSpectrosc.21352013.doi10.1255jnirs.10342.J.DahlbackaT.LillhongaandM.DringDesignedorthogonalsamplespikingbasedcalibrationsforquantitativeliquidphasemeasurementswithnearinfraredspectroscopyinananaerobicdigestionprocessJ.NearInfraredSpectrosc.21112013.doi10.1255jnirs.10323.J.DahlbackaandT.LillhongaQuantitativemeasurementsofanaerobicdigestionprocessparametersusingnearinfraredspectroscopyandlocalcalibrationmodelsJ.NearInfraredSpectrosc.21232013.doi10.1255jnirs.1033 JOURNALOFNEARINFRAREDSPECTROSCOPY1ISSN0967-0335IMPublicationsLLP2013doi10.1255jnirs.1037AllrightsreservedOn-linepredictionandcontrolanalysisofmanufacturingprocessesarestronglyrecommendedtomaintainandensurethequalityofnalproducts.Off-linemethodstomonitormanu-facturingprocessesareclassicallybasedonmanualsamplingoftenrequiringmultipledisruptionsduringtheprocessandresultinginsignicantsamplelossesandcostincreases.Inthepharmaceuticalindustrytheconceptofprocessanalyticaltech-nologyPATwasintroducedandinthepastdecadePATsystemsbasedonspectroscopictechniqueshavebeendeveloped.13Non-destructivenearinfraredNIRspectroscopyiswidelyusedtomonitormanufacturingprocessesincombinationwithchem-ometricsanalyticaltechniques.48Inpharmaceuticalprocessessamplehomogeneity135910moisture811crystallinepolymor-phism1214andpropertiesofnalproducts71517arekeyfeaturesrequiringreliableandconstantmonitoring.Typicallytwoblendingphasesarecarriedouttoobtainadirectcompressionformulation.Thepre-blendingphaseaimsatestablishingpowderhomogeneityintheabsenceoflubricantandisfollowedbythepost-blendingphasewhichinvolvestheadditionofalubricant.Inthepre-blendingphaseoneofthemostaccuratemethodstocharacterisehomogeneityisbasedontheuseofchemomet-ricstechniquessuchasthepartialleast-squaresPLSregres-sionmethod.Inthismethodtheestablishmentofaregressionmodelforeachcomponentisrequiredtopredictitsquantityandregressionmodelsarethuskeytoobtainingacceptablepredictions.Generallyinordertoestablishoptimisedmodelsanumberofreferencesamplesarerequired.Howeverinmostcasesmorethanthreecomponentsareincludedinpharma-ceuticalformulationswhichwouldthusleadtotheprepara-TabletcharacteristicspredictionbypowderblendingprocessanalysisbasedonnearinfraredspectroscopyYusukeHattoriYoshiyukiTajiriandMakotoOtsukaResearchInstituteofPharmaceuticalSciencesFacultyofPharmacyMusashinoUniversity1-1-20Shin-machiNishitokyo-shiTokyo202-8585Japan.E-mailmotsukamusashino-u.ac.jpPredictionofpowderblendhomogeneityandpharmaceuticalcharacteristicssuchaspowderowabilitymechanicalstrengthanddisintegrationtimeofcompressedsoliddosageswascarriedoutviaon-linenearinfraredspectroscopymeasurements.Blendhomogeneitywasevaluatedinthepre-blendingphaseusingbothpartialleast-squaresPLSandmodel-freeregressionmethods.Covarianceanalysisbetweenconsecutivespectrawasshowntorepresentafaithfulmonitoringofhomogeneityasamodel-freemethodwhichcoincidedwellwiththeresultsobtainedfromthePLSregressionmethod.Similarlyduringthepost-blendingphasewithlubricantpharmaceuticalpropertieswereaccuratelyandpreciselypredictedusingPLSregressionmodels.Thesepredictionmodelscouldbeimplementedinthedeterminationofoptimalblendingtimeforalubricantinviewofimprovingowabilityandofensuringadequatemechanicalstrengthandtabletdisintegrationtime.KeywordsnearinfraredspectroscopynearinfraredNIRwirelessspectrometerpowderblendingV-blendertabletcharacteristicschemometricspredictionmodel-freepredictionIntroductionY.HattoriY.TajiriandM.OtsukaJ.NearInfraredSpectrosc.21192013Received29August2012Revised3December2012Accepted10December2012Publication24January2013 2TabletCharacterisationbyPowderBlendingProcessAnalysistionofalargenumberofandinsomecasescostlyreferencesamples.Maesschalcketal.18reportedacharacterisationmethodforpowderblendingbasedonthescoreplotforthersttwoprincipalcomponentsPCswhileSekulicetal.2proposedamoving-blockstandarddeviationSDmethodusingNIRspectroscopyforthecharacterisationofhomogeneityinpharmaceuticalblends.Inthelattermethodmodelswerenotrequiredandintegrationofstandarddeviationsbetweenthreeconsecutivespectrawasimplementedforreal-timecharacterisation.Thusmodel-freepredictioncanbemoreeffectiveforactualproductionprocessesthanmodel-dependentmethodstopredictthehomogeneityofpowderblends.Inthisstudyseveralmodel-freemethodsincludingthemoving-blockSDmethodtopredictblendhomogeneitywerecarriedoutandthepredictedhomogeneitywasvalidatedwiththePLS-basedmodel-dependentmethod.Ontheotherhandduringthepost-blendingphaseinadequateorexcessivepost-blendingmayseverelyaffectproductqualityincludingnotonlyhomogeneitybutalsomechanicalstrengthdisintegrationanddissolutionpropertiesoftablets.1921Thesetabletpropertieswillinturnaffectbioavailability.Forexampleitiswelldescribedthatthelubricantmagnesiumstearatechangespropertiesbycoatingtheparticlesurface.21Withtheincreaseinsupplyanddemandofgenericmedicinebothreductioninproduc-tioncostsandensuringproductqualityarenecessaryforpharmaceuticalindustries.Sincethechangesintheproper-tiesofnalproductsareamatterofthegreatestsignicancethereal-timemonitoringofthepropertiesisdesiredtoensurethebioavailabilityofproducts.NIRspectroscopyrelatesnotonlytoquantitativebutalsoqualitativeinformationinvolvingmolecularinteractionsuchaswithhydrogenbonds.Generallyanumberofhydratedwaterandfunctionalgroupsincludingprotonssuchashydroxylandcarboxylgroupsexistinpharmaceuticalformulations.TheinteractionbetweenthesefunctionalgroupscontributestotheparticlebindingthusitshouldbepossibletopredictthetabletpropertiesandthelubricanteffectbymonitoringtheNIRspectralchangeinthemolecularinteraction.Hencethereal-timepredictionoftabletpropertieswasimplementedinthepost-blendingphaseusingNIRspectroscopyandPLSregressionmethods.AV-blenderiscommonlyusedtoblendpharmaceuticalpowders.Sekulicetal.2andEl-Hagrasyandcolleagues59usedabre-opticprobetocollectNIRspectraofpowdersinaV-blenderwiththeiroriginalsystems.InthisstudyawirelessdispersiveNIRspectrophotometerwasmountedontoaV-blendertomonitorthepowderblendinareal-timemanner.Sincethespectrometercanbeequippedwithanaccelerationsensortorecogniseitspositionduringrota-tionsspectracanbeefcientlycollectedwhenthespec-trometerrotatesby180fromthetopoftherotation.Inaddi-tionahighlysensitivecooleddiodearraydetectormakesitpossibletocollectspectrawithatemporalresolutionintheorderofmilliseconds.InthisstudybycombiningaV-blendertoawirelessNIRspectrometertheblendhomogeneityinthepre-blendingphasewascharacterisedusingbothPLSregressionmodel-dependentandmodel-freemethodsinordertovalidatethemodel-freemethod.Withtheadditionofalubricantblendowabilityandcharacteristicsofcompressedsoliddosageswerecalibratedandpredictedasafunctionofthepost-blendingtimeusingPLSregression.MaterialsandmethodsMaterialsModelformulationsoftheophyllinelactoseSuperTab21ANpharmacopoeialpotatostarchandmagnesiumstearateMgStwereusedtodemonstratethemonitoringofhomogeneityandthepredictionofformulationcharacteristics.TheophyllinewaspurchasedfromSizuokacaffeineSizuokaJapan.Supertab21ANwasobtainedfromDMV-FonterraexcipientsVeghelNetherland.PotatostarchandMgStwerepurchasedfromKozakaiSeiyakuTokyoJapanandWakoPureChemicalsTokyoJapanrespectively.Allmaterialsweresievedthroughan850mmeshpriortouse.MethodsNIRspectralacquisitionNIRspectraforbothcalibrationandpredictionpurposeswerecollectedusingawirelessNIRspectrometerLancIRIIBrukerOpticsEttlingenGermany.Thedimensionsofthespectro-metermodulewere280mm225mm140mmwdhand7.2kginweight.TheLancIRIIwasconnectedtoacomputerviaanIEEE802.11-bstandardwithWEP128-bitencryptionwithina2metredistance.NIRspectrawereacquiredwith50accumulationsand87sexposuretime.Theacquiredspectrarangedwithin1100nmand2200nmwitha15nmresolutionand1nmwavelengthaccuracy.Thedetectorwascomposedofa256-elementTE-cooledInGaAsdiodearray.Theavailablesamplingareaandworkingdistancewere25mmindiameterand35mmrespectively.AglasswindowwasinstalledinthecapoftheV-blendertoefcientlyacquirespectra.TheLancIRIIwasmountedonthecapwithanoriginaljunction.TheLancIRIIwasequippedwithanaccelerationsensorthusintheblendingprocedurespectralacquisitionwastrig-geredbyitsrotationalposition.Bydeninganangleof0astheupsideoftheblendingsystemspectrawereacquiredbetween160and190rotationangle.BlendingprocedureSievedpowderswereloadedintoa5LV-blenderintheorderoflactosestarchandtheophylline.Theamountofeachmaterialusedwas1248g63534g27and198g10respec-tively.Thesepowderswereblendedtobeuniformat20rpmrotationaheadoftheadditionofMgSt.Thepre-blendingwascarriedoutfor5minwithNIRspectralmonitoring.MgStwasaddedintotheblendedpowderandtheblenderwasactivatedagainat20rpmrotationforafurther120min. Y.HattoriY.TajiriandM.OtsukaJ.NearInfraredSpectrosc.211920133TheprocessofMgStblendingwasalsomonitoredbyNIRspectralmeasurements.Threedistinctbatchesofpowderblendingwereperformedandmanualsamplingduringpost-blendingwasperformedontherstbatch.CalibrationbynearinfraredspectroscopyCalibrationsofblendratiosandothercharacteristicsi.e.angleofreposeofblendshardnessanddisintegrationtimeoftabletsweremodelledbasedonNIRspectraofpowderblendsusingthePLSregressionmethod.PLSregressionwasperformedusingchemometricssoftwarePirouette3.11InfometrixBothellWAUSA.Blendratioswerecalibratedusing18samplesblendedbyhand.Thepreparedsampleswereheapedontothefenes-tratedcapandtappeddownbeforecollectingNIRspectra.Thespectrawerecollectedthreetimespersampleusingdifferentacquisitionpositions.PLSregressionwasperformedusingspectrawhichhadbeenpre-processedwithmultiplicativescatteringMSCcorrelationandmean-centring.Threeorfourlatentvariableswereoptimalandleave-three-outstep-validationwasimplementedasaregression.DuringtheblendingofMgStpowderblendsweresampledfromtheblenderatvedistinctpositionstocalibratethechar-acteristics.Themanualsamplingwasperformedateighttimepointsof5min10min20min30min45min60min90minand120minpast.NIRspectrawereacquiredateveryrotationviatheLancirIIfor120min.Powderowabilitywasevaluatedbymeasuringtheangleofrepose.Inadditionthesampledpowder200mgwascompressedwithapressureof279MPausingacompressorTG5-KNMinebeaNaganoJapanandat-surfacepunchesf8mm.Themechanicalstrengthofthetabletswasmeasuredusingadigitalload-celltypehardnesstesterPortableCheckerPC-30OkadaSeikoTokyoJapan.TabletdisintegrationtestswereperformedusingatabletdisintegrationapparatusNV-2FToyamaSangyoOsakaJapan.Thetabletdisintegrationtimewasvisuallymeasuredindistilledwaterwithadiscatatemperatureof37C2C.Averagetimeandstandarddeviationswereobtainedwithmeasurementsfromthreetabletspersample.ThesecharacteristicsweremodelledbyPLSregressionbasedontheNIRspectraandmeasuredcharacteristicsatdistincttimepoints.TheusedNIRspectrawerepre-processedwithMSCandmean-centringandleave-one-outcross-validationwasappliedforPLSregressionwithtwolatentvariables.ResultsanddiscussionHomogeneityinpre-blendingRegressioncurvesforblendratiosofeachpowdercomponentweremodelledusingpreparedsamples.TheNIRspectraofthesamplesareshowninFigure1.AbsorptionbandsduetocombinationOHstr.COHbend.andrstovertoneOHstr.modesofbothstarchandlactosewerefoundataround2100nmand1580nmrespectively.Thebandsataround1930nm1470nmand1360nmwereassignedtocombinationmodesofwater1940nmHOHbendingandOHasym-metricstretching1470nmand1360nmOHsymmetricandasymmetricstretchingwhichadsorbedtostarchviahydrogenbonds.Weakbandsataround1680nmwereduetorstover-toneofCHstretching.Bandsdueto2ndovertoneofCHstretchingwerefoundataround1200nm.22WavelengthnmAbsorbanceab2100nm1930nm1680nm1580nm1470nm1200nm1360nm120015001800210000.10.20.30.4Figure1.aNIRspectraofpreparedsamplesforquantitativecalibrationsandbrepresentationoftheirsecondderivatives.MeasuredblendratiowtPredictedblendratiolactosestarchtheophyllinewt0204060204060Figure2.QuantitativecompositioncalibrationlinesoflactoseflactosegreytrianglesstarchfstarchclosedtrianglesandtheophyllineftheophyllineopencirclesusingthePLSregressionmethod.ThesePLSregressionswereoptimisedbytwolatentvariablesusingthewholespectralrange. 4TabletCharacterisationbyPowderBlendingProcessAnalysisFigure2isrepresentedregressionlinesforeachcomponentwhichshowednelinearlineswhichcrossedtheoriginandtheslopeswithequaltonearly1.Table1showstheresultsoferroranalysisintheregressionmodelswherestandarderrorofpredictionSEPwascalculatedfrompredictionresidualerrorsumofsquaresPRESSasthefollowingequation1PRESSSEPn212-nxxPRESSfxfwherefxfxandnindicatetheactualdatasetforthevali-dationpredicteddatasetandthenumberofvalidationdatarespectively.RegressionvectorsfortheseregressionmodelsareshownwiththeNIRspectrumofeachpurecomponentinFigure3.Severaloverlapscanbeobservedbetweenregressionvectorsandcomponentspectra.BandsduetocombinationofOHstretchingandCOHbendingmodesandtherstovertoneofOHstretchingcontributedtothemodel-lingoflactosequantitativecalibration.Theregressionvectorofstarchindicatesthatthecalibrationwasmostlycontributedbythebandsduetohydratedwater.ThelowestcontentoftheophyllinewasnelymodelledwiththebandsofrstandsecondovertonesofCHstretching.Thesecalibrationswerebasedondifferentabsorptionbandsforeachcomponentthusthesemodelscanprovidequantitativepredictionofcomponentsindependentlyofeachother.HomogeneityprolesofeachcomponentwerepredictedusingthemodelsasafunctionofrotationsoftheV-blenderasshowninFigure4.Thehomogeneitywasdefinedastheratioofweightpercenttothetheoreticalcontent.Themostflowableanhydrouslactosequicklyapproachedthetheoreticalcontentandwithin15rotationsthehomogeneitywasmorethan95andheldsteadyataround97.Potatostarchapproached95ofthetheoreticalconcentrationwithinapproximately30rotationsandreachedasteadystatefollowinganadditional30rotations.Theconcentra-tionoftheophyllineslowlydecreasedto105ofthetheo-reticalvaluewithin60rotations.Morethan80rotationsjlactosejstarchjtheophyllineARaMSbDTcSlope0.9970.9950.9980.9690.9580.706Intercept0.1720.1100.06061.343.48N89.7sRcv20.9960.9960.9980.9680.9580.701SECV0.5520.3440.5970.444.36N25.9sNFd443323aangleofreposebmechanicalstrengthcdisintegrationtimednumberoffactors.Table1.Erroranalysisforthecalibrationresultsofpartialleast-squaresPLSregression.Figure3.RegressionvectorsforcompositioncalibrationssolidlinesandpureNIRspectrabrokenlinesofalactosebstarchandctheophylline.00.20.4-1001000.20.4-10010WavelengthnmAbsorbanceRegressionvectorABC120015001800210000.2-10010abcNumberofrotationscountsPredictedhomogeneity020406080100708090100110120130Figure4.Quantitativepredictionsoflactosegreytrianglesstarchclosedtrianglesandtheophyllineopencirclesasafunctionofthenumberofrotationsoftheblender.Thehomogeneitywascharacterisedastheratioofpredictedtothetheoreticalcontentinpercentage.Dottedandsolidlinesindicate90and95oftheoreticalcontentrespectively. Y.HattoriY.TajiriandM.OtsukaJ.NearInfraredSpectrosc.211920135wererequiredtoreachastationarystateatwithin102ofthetheoreticalvalue.Thesehomogeneityproleswerenotexpectedtobe100becausepowderblendsareessentiallyheterogeneous.Model-freecharacterisationsofpowderblendhomogeneitywerealsodemonstratedwithtwoalgorithms.Therstwasmoving-blockSDwithveastheblocksize.HomogeneitywascharacterisedbytheintegrationofSDateachwavelengthwithveconsecutivespectradenedasi2i1ii1andi2thustheithintegrationwasrepresentativeofthemall.Thesecondwasbasedontheuseofcorrelationcoefcientandcovarianceanalysisfromonespectrumtothenext.IntegrationofSDwasplottedasafunctionofrotationinFigure5a.Theintegrationwasreducedtonearly0within15rotationswhichwasasfastasthehomogenisationoflactose.LowSDindicateshighsimilaritybetweenthefivespectrathustheblendhomogeneitywascharacterisedtoreachasteadystatewithinafewcountsofrotation.ProlesofcorrelationcoefcientandcovariancebetweentwoconsecutivespectraareshowninFigure5b.Thecorrela-tioncoefcientquicklyapproached1whichsuggestshighposi-tivecorrelationbetweentwospectrahoweverthecovariancegraduallyincreasedasthecorrelationcoefcientreachedastationarystate.Figure5calsoshowsprolesofcorrelationcoefcientandcovariancecalculatedusingsecondderivativesofNIRspectra.Thecorrelationcoefcientapproached0.999assoonasisshowninFigure5b.Thecovarianceincreasedslowlyandmorecontinuouslycomparedtothecovarianceprolewithoutsecondderivation.Thecovarianceseemedtoattainastationarystateat7080rotations.Thehomogeneityofmostflowableanhydrouslactosequicklyreachedauniformconditionandwasstationaryunderthesameconditions.Thefastblendingwascontributednotonlybythegoodowabilitybutalsothemajorcomponentoflactose.Sincepotatostarchismadeofcoherentparticlesitsowabilityispoorandthusblendingisslow.Theophyllineexhibitedaslowprogressiontowardsauniformstateandhadwideuctuationduetoitbeingtheminorcomponentintheblend.ObviouslyprogresstowardshomogeneityvariedbetweenthecomponentsthustheprogressinintegratedhomogeneityUwasobtainedasasumoftheweightedcontentofcomponentsasfollows3jjjjj11001003632710theophyllineNilactosestarchiiUNwhereNrepresentsthenumberofcomponentsjrepresentsthecontentofacomponentandthesubscriptindicatesthetheoreticalcontent.ThecalculatedprogressesofallbatchesisshowninFigure6.Thehomogeneitywasprogressivelyimprovedandreachedastationarystatewithintheuctua-tionsof2.Essentiallyapowderblendisheterogeneousanddoesnotbecomecompletelyhomogeneousthustheuc-tuationreectsactualheterogeneityoftheblend.Basedontheregressionmodel-dependentmethodthepowderblendsweredeterminedtobehomogenisedwithin60rotations.Ontheotherhandhomogeneitycharacterisationviamodel-freemethodswasdemonstratedinFigure5.CorrelationcoefficientsdescribedafastprogressiontoaFigure5.Homogeneitycharacterisationsbyamodel-freemethodsofintegrationofstandarddeviationSDbcorrelationcoefcientdiagonalcrossesandccovarianceopencirclesobtainedfromthecollectedNIRspectraandthesecondderivatives.Thesecharacterisationmethodswereimplementedusingthesamespectraldatasetastheoneusedintheregressionmodel-dependentmethodshowninFigure4.IntegratedhomogeneityUNumberofrotationscountsNormalizedcovariance105030609010012014000.51Figure6.IntegratedhomogeneitycharacterisationusingEquation3opensymbolsandnormalisedcovarianceprolesclosedsymbols.Therstsecondandthirdbatcheswererepresentedbycirclestrianglesandrectanglesrespectively.Covariancewasnormalisedtomaximumandminimum.CovarianceCovarianceCorr.Coeff.Corr.Coeff.0.01250.0130.970.980.991Integrationofstandarddeviation00.050.10.150.2ABCRotationcounts0204060801001.61.71.810-90.90.951abc 6TabletCharacterisationbyPowderBlendingProcessAnalysisstationarystate.Inthisalgorithmspectraduetothemajorcomponentoflactoseeffectivelycontributedtotherepresen-tationthustheresultwasnotapplicabletocharacterisationofthetotalhomogeneity.Themoving-blockSDprovidedadistincthomogeneityproletothatpredictedbytheregressionmodel.Sekulicetal.andotherauthorsreportedfastprogressiontohomo-geneityusingthemoving-blockSDmethod.291023InourstudySDwaslargelydeterminedbythemajorcomponentoflactoseandtheothercomponentsmademinorcontributionstotheprogress.Ontheotherhandcovariancebetweentwoconsecutivespectrarepresentedagradualprogression.MoreparticularlythesecondderivativeprovidedacontinuousproleinallthreebatchesasshowninFigure6.Covarianceincreaseduntil60rotationsandbecamealmostconstantbeyond60rotations.Howeveritisdifculttocharacterisethehomogeneityquanti-tativelyusingcovariancebecausetheabsolutevaluedependsontheintensityoftheoriginalspectrum.Comparisonbetweenregressionmodel-dependentandmodel-freemethodsindicatesthatbothresultscoincidedwellwithpowderblendhomogeneity.Theregressionmodel-dependentmethodinvolvesanapplicablelimitationandmoreoverthepreparationandvalidationofregressionmodelsforvariousblendsystemsincurshugecostsinmanycases.Thereforeweproposethatthemodel-freemethodbasedoncovarianceanalysisbetweenspectrawhichprovidescoin-cidentpredictionwithmodel-dependentmethodsismoreviableandeconomicallyadvancedthantheothermethods.Tabletcharacteristicspredictioninpost-blendingphaseFollowingapre-blendinglubricantofMgStwasaddedintoaV-blenderandthepowderwasfurtherblendedat20rpmwithNIRspectralmonitoring.Duringtherstruntherota-tionoftheblenderwashaltedsothatpowdersamplescouldbecollectedat5min10min20min30min45min60min90minand120minaftertheadditionofMgSt.RegressionsforangleofreposemechanicalstrengthanddisintegrationtimeweremodelledbycombiningactualmeasurementsoftheirpropertiesandmonitoredNIRspectra.TheresultsofregressionandtheerroranalysisareshowninFigure7andTable1respectively.Decreaseintheangleofreposeindicatesimprovementinpowderowabilityuponblending.Althoughthemechanicalstrengthofthetabletsdecreasedafter60minofblendingthedisintegrationtimedecreasedat10minbutwaselongatedafter30minofblending.LoadingsofLV1andLV2ABC1360nm-0.2-0.100.10.2-0.2-0.100.10.2Wavelengthnm1200150018002100-0.2-0.100.10.2bacFigure7.CharacteristicscalibrationlinesofaangleofreposebmechanicalstrengthandcdisintegrationtimebyPLSregressionwiththreelatentvariables.Numbersingraphsindicatepointsintimemin.ofmanuallysampledfromtheblender.5102030456090120MeasuredangleofreposePredictedangleA4044484044485102030456090120MeasureddisintegrationtimesPredictedtimesC250300350400250300350400PredictedstrengthNMeasuredmechanicalstrengthN5102030456090120B60801001206080100120abcFigure8.LoadingsoftherstandsecondlatentvariablesLV1andLV2forcharacteristiccalibrationsofaangleofreposebmechanicalstrengthandcdisintegrationtime.BlackandgreysolidlinesindicateloadingsofLV1andLV2respectively.BrokenlinesindicatetheNIRspectrumofstarch. Y.HattoriY.TajiriandM.OtsukaJ.NearInfraredSpectrosc.211920137Theseregressionmodelsshowednelinearcurvescrossingtheoriginwithaslopeofnearly1.Figure8showstheloadingsoffirstandsecondlatentvariablesfortheirmodelsandNIRspectrumofpotatostarch.Starchisnotveryowablebuthasgoodcompressionmold-abilityandisusedasadisintegrantthusthecontributionsofstarchtotheirpropertiesarecomparablyhigherthantheothercomponents.Theloadingshadlowqualitysignal-to-noiseratiosbecauseoftheminimalchangeinspectradependingontheproperties.IntheloadingsofPC1peakswereobservedintheranges13001450nm1450nm1650nmand18001950nm.IntheloadingsofPC2asharppeakwasfoundat1360nmandbroadpeakswereobservedintheranges14501650nmand18802050nm.Thepeaksintheranges18802050nmand14501650nmwereduetoacombinationmodebetweenOHstretchingandHOHbendingofhydratedwaterandrstovertoneofCOHstretchingrespectively.22Thepeaksinthebothranges13001450nmand18001950nmofPC1loadingswereassignedtothebandsduetofreeorweaklyhydrogen-bondedwater24whichwereenhancedbytheblendingofMgSt.ThustheinteractionbetweenMgStandstarchinducesdehydrationofstarchandasaconsequencetheowabilityofstarchisimprovedanditsagglutinationprop-ertyinhibited.TheregressionmodelswereappliedforthecontinuouspredictionoftheirpropertiesfromNIRspectralmeasure-ments.ThepredictedresultsobtainedfromthreebatchesareshownwithactualmeasurementsrstbatchinFigure9.Onlyonebatchthirdbatchindicatedfasterchangesintheircharacteristicscomparedtotheothertwobatcheshowevertheirmodelsworkedsuccessfullyandtheactualmeasure-mentscoincidedwellwiththepredictedvalues.Theangleofreposedecreaseduponblendingandreachedsaturationafter60minthusindicatingthatthepowderstruc-turebecameanorderedmixtureaftera60minblendingperiod.Mechanicalstrengthwasattenuatedupon60minofblendingsincetheorderedmixturedisturbedcohesionbetweenparti-clesandimprovedowability.AtthebeginningofMgStblendingthehardtabletsdidnoteasilydisintegrateinwaterbutasubsequentdecreaseinmechanicalstrengthprecipitatedthedisintegrationofthetablets.HoweveruponincreasedinteractionsbetweenMgStandstarchparticlesastheblendingwentonaslowdisintegrationwasinducedafter30minofblending.SincethestearicacidchainsofMgStactasasurfactantweproposethattheMgStchangesthesurfacehydrophobicityofstarchduetothealkylchainsofstearin.25Thehydrophobiclayerofstearicaciddisturbedthewaterpermeationintothedisintegrantofstarchandaccordinglydisintegrationofthetabletswasdelayed.ThepredictedprolesshowninFigure9provideinforma-tionontheoptimalblendingtimeofMgSttoobtainadequateflowabilitysufficientmechanicalstrengthandmoderatedisintegrationtime.Althoughintherstandsecondbatchestheircharacteristicswereoptimisedata43angleofrepose80Nofstrengthand280sofdisintegrationtimeatapproxi-mately30minofblendinginthe3rdbatchfasterchangesintheircharacteristicswerepredictedanditwasdeterminedthata15minblendingperiodresultedinthemostappro-priatecharacteristics.Inadequateblendingcausedpoorowabilityandexcessiveblendinginducedlowerstrengthandslowerdisintegrationoftablets.Thereal-timeprocessanal-ysisprovidesafeedbacktocontrolmanufacturingprocessesefciently.ConclusionInthepresentstudywesuccessfullydemonstratedthereal-timepredictionofpowderhomogeneityandotherpharma-ceuticalpropertiesusingchemometricandnon-chemo-metricNIRspectroscopy.Thehomogeneityofanhydrouslactosepotatostarchandtheophyllinewerewellcharacter-isedusingPLSregressionmodels.Inadditioncovarianceanalysisbetweentwoconsecutivespectrawasusedtochar-acterisethehomogeneityqualitativelywithoutregressionmodels.Bycomparingregressionmodel-dependentandmodel-freemethodsweshowedthatquantitativelyequiv-alentresultswereobtainedinthefirstderivativesoftheFigure9.ActualmeasurementsopencirclesandpredicteddiagonalcrossesaangleofreposebmechanicalstrengthandcdisintegrationtimeinthreebatchesasafunctionofblendingtimeofMgSt.Thecoloursofcrosssymbolsindicaterstblackseconddarkgreyandthirdlightgreybatches.ABC404550StrengthN50100150AngleofreposeDisintegrationtimesBlendingtimeofMgStmin0306090120200300400bac 8TabletCharacterisationbyPowderBlendingProcessAnalysisprolesthuswesuggestthatthemodel-freecharacterisa-tionofhomogeneityisappropriateandmoreeffectivethanthemodel-dependentmethod.Duringthepost-blendingwithMgStangleofreposeofblendedpowdermechanicalstrengthanddisintegrationtimeofcompressedsoliddosageswereaccuratelyandpreciselypredictedbythePLSregressionmethodandthepredictionscoincidedwithactualmeasurements.Inadequateblendingcausedpoorflowabilityandlowerstrengthwithslowerdisintegrationoftabletscausedbyexcessiveblending.Themostsignicantpointforthepharmaceuticalindustryistoensurestablequalityofproducts.Moreparticularlytabletpropertiesdirectlycontributetothebioavailabilitythusthereal-timepredictionofthepropertiesduringthepost-blendingphasehelpstoimprovethequalityandprovidessubstantialbenetstotheindustries.References1.W.PluggeandC.vanderVliestTheuseofnearinfraredspectroscopyinthequalitycontrollaboratoryofthepharmaceuticalindustryJ.Pharm.Biomed.Anal.1010127971992.doi10.10160731-70859180083-L2.S.S.SekulicH.W.WardD.R.BranneganE.D.StanleyC.L.EvansS.T.SciavolinoP.A.HaileyandP.K.AldridgeOn-linemonitoringofpowderblendhomogeneitybynear-infraredspectroscopyAnal.Chem.6835091996.doi10.1021ac962024n3.D.J.WargoandJ.K.DrennenNear-infraredspectroscopiccharacterizationofpharmaceuticalpowderblendsJ.Pharm.Biomed.Anal.141114151996.doi10.10160731-70859601739-64.O.BerntssonL.-G.DanielssonB.LagerholmandS.FolestadQuantitativein-linemonitoringofpowderblendingbynearinfraredreectionspectroscopyPowderTechnol.1231852002.doi10.1016S0032-59100100456-95.A.S.El-HagrasyandJ.K.DrennenAprocessanalyticaltechnologyapproachtonear-infraredprocesscontrolofpharmaceuticalpowderblending.partIIIquantitativenear-infraredcalibrationforpredictionofblendhomo-geneityandcharacterizationofpowdermixingkineticsJ.Pharmacol.Sci.9542220056.M.OtsukaY.MouriandY.MatsudaChemometricevaluationofpharmaceuticalpropertiesofantipyrinegranulesbynear-infraredspectroscopyA.A.P.S.PharmSciTech.431422003.doi10.1208pt0403477.M.OtsukaandI.YamanePredictionoftablethardnessbasedonnearinfraredspectraofrawmixedpowdersbychemometricsJ.Pharmacol.Sci.95714252006.doi10.1002jps.205148.J.RantanenH.WikstrmR.TurnerandL.S.TaylorUseofin-linenear-infraredspectroscopyincombinationwithchemometricsforimprovedunderstandingofpharmaceuticalprocessesAnal.Chem.7725562004.9.A.S.El-HagrasyH.R.MorrisF.DAmicoR.A.LodderandJ.K.DrennenAprocessanalyticaltechnologyapproachtonear-infraredprocesscon-trolofpharmaceuticalpowderblendingpartIIqualitativenear-infraredmodelsforpredictionofblendhomogeneityJ.Pharmacol.Sci.90212982001.doi10.1002jps.108210.M.BlancoR.G.BaandE.BertranMonitoringpowderblendinginpharmaceuticalprocessesbyuseofnearinfraredspectroscopyTalanta5612032002.doi10.1016S0039-91400100559-811.Y.RoggoN.JentA.EdmondP.ChalusandM.UlmschneiderCharacterizingprocesseffectsonpharmaceuticalsolidformsusingnear-infraredspec-troscopyandinfraredimagingEur.J.Pharm.Biopharm.61121002005.doi10.1016j.ejpb.2005.04.00512.G.BucktonE.YonemochiJ.HammondandA.MoffatTheuseofnearinfra-redspectroscopytodetectchangesintheformofamorphousandcrystallinelactoseInt.J.Pharm.1682311998.doi10.1016S0378-51739800095-713.K.KamadaS.YoshimuraM.MurataH.MurataH.NagaiH.UshioandK.TeradaCharacterizationandmonitoringofpseudo-polymorphsinmanufacturingprocessbyNIRInt.J.Pharm.368121032009.doi10.1016j.ijpharm.2008.10.01014.M.BlancoD.ValdsI.LlorenteandM.BayodApplicationofNIRspectroscopyinpolymorphicanaly-sisstudyofpseudo-polymorphsstabilityJ.Pharmacol.Sci.94613362005.doi10.1002jps.2036215.M.DonosoandE.S.GhalyPredictionofdrugdissolutionfromtabletsusingnear-infrareddiffusereectancespectroscopyasanondestructivemethodPharm.Dev.Technol.92472005.doi10.1081PDT-20003142316.M.DonosoD.O.KildsigandE.S.GhalyPredictionoftablethardnessandporosityusingnear-infrareddiffusereectancespectroscopyasanondestructivemethodPharm.Dev.Technol.843572003.doi10.1081PDT-12002468917.M.OtsukaH.TanabeK.OsakiK.OtsukaandY.OzakiChemoinformetricalevaluationofdissolutionpropertyofindomethacintabletsbynear-infraredspectros-copyJ.Pharmacol.Sci.9647882007.doi10.1002jps.2070418.R.DeMaesschalckF.C.SanchezD.L.MassartP.DohertyandP.HaileyOn-linemonitoringofpowderblendingwithnear-infraredspectroscopyAppl.Spectrosc.527251998.doi10.1366000370298194414819.G.K.BolhuisA.J.SmallenbroekandC.F.LerkInteractionoftabletdisintegrantsandmagnesiumstearateduringmixingIeffectontabletdisintegrationJ.Pharmacol.Sci.701213281981.doi10.1002jps.2600701210 Y.HattoriY.TajiriandM.OtsukaJ.NearInfraredSpectrosc.21192013920.D.S.DesaiB.A.RubitskiS.A.VariaandA.W.NewmanPhysicalinteractionsofmagnesiumstearatewithstarch-deriveddisintegrantsandtheireffectsoncap-suleandtabletdissolutionInt.J.Pharm.912171993.doi10.10160378-51739390341-C21.H.AbeandM.OtsukaEffectsoflubricant-mixingtimeonprolongationofdissolutiontimeanditspredictionbymeasuringnearinfraredspectrafromtabletsDrugDev.Ind.Pharm.3844122011.22.P.WilliamsandK.NorrisEdsNearInfraredTechnologyintheAgriculturalandFoodIndustriesAmericanAssociationofCerealChemistsStPaulMinneapolisUSA1987.23.P.A.HaileyP.DohertyP.TapsellT.OliverandP.K.AldridgeAutomatedsystemfortheon-linemonitoringofpowderblendingprocessesusingnear-infraredspec-troscopypartI.SystemdevelopmentandcontrolJ.Pharm.Biomed.Anal.1455511996.doi10.10160731-70859501674-024.V.FornesandJ.ChaussidonAninterpretationoftheevolutionwithtemperatureofthen2n3combinationbandinwaterJ.Chem.Phys.6846671978.doi10.10631.43557625.P.BracconiC.AndrsandA.NdiayeStructuralprop-ertiesofmagnesiumstearatepseudopolymorphseffectoftemperatureInt.J.Pharm.262121092003.doi10.1016S0378-51730300339-9 J.DahlbackaT.LillhongaandM.DringJ.NearInfraredSpectrosc.2111222013Received17April2012Revised25September2012Accepted28October2012Publication21January2013ISSN0967-0335IMPublicationsLLP2013doi10.1255jnirs.1032Allrightsreserved11JOURNALOFNEARINFRAREDSPECTROSCOPYBiogasproductionthroughanaerobicdigestionADplaysanimportantroleintheefforttoreducetheamountofgreenhousegasemis-sionsfromenergyproduction.Atthesametimecommercialdigestionprocessesarecommonlyoperatedwellbelowtheirmaximalcapacity.Onereasonforthisisthelackofsuitablesensorsformonitoringkeyparametersintheprocess.NearinfraredNIRspectro-scopyisseenasapotentiallyinterestingtechniqueforliquidphasemeasurementsinADprocesses.InthisworktheconcentrationsoftheimportantconstituentsammoniumacetatepropionateandtotalvolatilefattyacidsTVFAweremeasuredbycombiningNIRtransmittancemeasurementsandpartialleast-squaresPLSmodels.Inordertoobtainadditionalcalibrationdatawithouthavingtoperformadditionalreferencemeasurementsandinordertoenhancetheconstituentspeciccorrelationinthecalibrationdataaspikingprocedurewasimplementedaccordingtoacentralcompositedesign.ThismethodologyreducedtherootmeansquareerrorofpredictionRMSEPforammoniumfrom176mgL1to127mgL1foracetatefrom334mgL1to260mgL1forpropionatefrom258mgL1to203mgL1andforTVFAfrom858mgL1to704mgL1.Thespikingprocedurealsosignicantlyincreasedthecorrelationbetweenmodelpredictionsonpureconstituentspectraandreferencevaluescomparedtothatofmodelsbasedonnon-spikeddata.IntroductionThecomplexprocessofmethanefermentationisexpectedtoplayavitalroleinthefutureasaversatilerenewableenergyresourceprovidingafuelalternativethatcandras-ticallydecreasetheemissionofgreenhousegasestotheatmospherebyreducingtheconsumptionoffossilfuelsandatthesametimeprovidingasubstituteformineralferti-lisersthroughthedigestate.1HowevercommercialanaerobicdigestionADprocessesarecommonlyfarfromoptimisedmainlyduetoalackofsuitablemeasurementequipmentthatwouldenablethemonitoringandcontrolofkeyparameters.2IthasbeensuggestedthatnearinfraredNIRspectroscopycouldplayanimportantroleinovercomingthisobstaclebyprovidingadetailedpictureofthemicrobialprocessesinthedigesters.3TheachievementsofNIRspectroscopyhavealreadybeendocumentedinanumberofpublicationsasmeritoriouslyreviewedbyMadsenetal.4ThepresentstudyaimstocontributetothealreadyexistingknowledgeabouttheusefulnessofNIRspectroscopyformonitoringpurposesofAD.Howevertheemphasisisputonamethodologyforproducingadequatecalibrationinformationratherthanonthemeasurementitselfortheimpactofthemeasurementforprocessoptimisationpurposes.KeywordsanaerobicdigestionsamplespikingdesignofexperimentsbioprocessmetabolitesNIRammoniumacetatepropionatecentralcompositedesignDesignedorthogonalsamplespikingbasedcalibrationsforquantitativeliquidphasemeasurementswithnearinfraredspectroscopyinananaerobicdigestionprocessJohnDahlbackaaTomLillhongaaandMarcoDringbaNoviaUniversityofAppliedSciencesPOBox6FIN65201VaasaFinland.E-mailJohn.Dahlbackanovia.bUniversityofWismarPOBox121023952WismarGermany 12OrthogonalSampleSpikingBasedCalibrationsforQuantitativeLiquidPhaseMeasurementsThisstudywasinitiatedinordertoevaluatetheusabilityofaportablediodearrayNIRinstrumentforquantitativeliquidphasemeasurementsofkeymetabolitesandintermedi-atesintheproductionofbiogaswithanaerobicdigestionperformedasbatchprocesses.Itwasassumedthatthiswouldbechallengingduetotherelativelylowconcentra-tionsofthekeyconstituentsandcorrespondinglyhighlevelsofwaterabsorbanceandthecomplexityofthebackgroundmatrixduetomicrobialactivityandtheoriginofthesubstrate.Furthermorethemultiplicativescatteringassociatedwithparticulatematterthelowopacityoftheliquidphasemakingtransmissionmeasurementsdifcultandthenaturalinter-correlationduetomicrobialactivitybetweenconstituentsofinterestandconstituentsinthebackgroundmatrixaswellwereseenaschallengesintheimplementationofthemeasurement.AnexampleofconstituentintercorrelationinADcanbefoundintheformofacorrelationmatrixbyHolm-Nielsen.5Thismatrixincorporatesthecorrelationbetween11constituentsinsamplesobtainedfromfull-scalemethanegasproductionunits.Disregardingthematrixdiagonalthematrixcontainsnineintercorrelationvaluesabove0.9andanadditionalfourabove0.8.Itwasthereforeassumedthatthesamplesobtainedfromlaboratoryreactorsoperatinginbatchmodewouldalsocontainintercorrelatedconstituents.Itwasalsoassumedthatitwouldbeofinteresttoevaluatehowthemeasurementwouldperformwithoutconstituentintercorrelation.Itwasdecidedthatthisevaluationwouldbeconductedbyspikingprocesssamplesandthatthespikingofthesampleswouldbecarriedoutaccordingtoacentralcompositedesign.Thetermspikingiscommonlytakentomeantheadditionofaknownquantityofanalytetoamatrixsimilaroridenticaltothesampleofinterest.6Thecentralcompositedesigncanbedescribedasanaturalextensionofthetwolevelsfullandfractionaldesigns.7Spikingofprocessbrothforcalibrationpurposesinbioprocessmonitoringusingvibrationspectroscopyisperhapsnotaparticularlynewideabutitappearsthatduringthelast20yearsithasbeenpractisedonlybyafew.812Someexampleswherespikinghasbeenusedinthesamecontextbutrathertoevaluatecalibrationmodelaccuracycanalsobefound.1314Howeveranyexamplesonusingmulti-constituentspikesforcalibrationmodelregressioninaccordancewithanorthogonaldesignwerenotfound.Itwasassumedthatinadditiontothereducedriskofcreatingcalibrationmodelsthatincludeintercorrelationeffectsthespikingproce-durecouldalsoserveasanexampleofhowtheamountofdataavailableformodelregressioncanbeincreasedwithoutrequiringadditionalsamplingfromthereactorsoradditionalreferencemeasurementsandthusalsoenablerelativelyfastgenerationofthecalibrationdata.Theselec-tionofconstituentsinthespikingdesignwasmainlybasedontheexpectedrelativelyhighconcentrationlevelsoftheselectedconstituentsinthesamplesincombinationwiththeconstraintthattheconstituentshouldbeamolecularlydenedconstituent.Thustheselectionbecameammoniumacetateandpropionate.InadditioncalibrationmodelswerealsoevaluatedfortotalvolatilefattyacidsTVFA.AlthoughtheconstituentswereselectedprimarilyfromaNIRspec-troscopicquantitativemeasurementsuitabilitycriterionitcanbepointedoutthattheselectedconstituentsarealsoofrelevanceasprocessvariables.ForinstancevolatilefattyacidsVFAareconsideredtobesomeofthemostimpor-tantintermediatesinbiogasproductionthroughanaerobicdigestion15whereasammoniaisnecessaryforbacterialgrowthatthesametimeasinorganicammonianitrogenalthoughfreeammoniahasbeensuggestedtobethemaincauseisawell-documentedinhibitorofmethanogenesis.16MaterialsandmethodsMastersolutionsForeachofthethreeconstituentsammoniumacetateandpropionateamastersolutionwithanionconcentra-tionof100gLsolution1wasprepared.Thesolutionswereobtainedbydissolvingsaltscontainingtheionofinterestinwater.Thesaltsusedforammoniumacetateandpropi-onatewereammoniumsulphatemagnesiumacetateandcalciumpropionate.Theamountofeachsaltneededtoobtain750mLofa100gL1solutionwascomputedandmeasuredandthesaltwasthereafterplacedinameasure-mentbottle.Approximately600mLdistilledwaterwasaddedtothebottleanditwasplacedonaheatedmagneticstirrer.Afterthesaltwasdeterminedtobecompletelydissolvedaccordingtoavisualassessmentadditionalwaterwasaddeduntilatotalvolumeof750mLwasobtained.Themastersolutionswerethenplacedinarefrigeratorforlateruse.PureconstituentsamplesForeachofthethreeconstituentsammoniumacetateandpropionatepureconstituentsolutionswithaconcentrationrangefrom0.5gL1to5gL1ionconcentrationpersolutionvolumewereprepared.Inadditiontothesaltsusedinthepreparationofthemastersolutionsammoniumchloridesodiumacetateandsodiumpropionatewerealsousedinthepreparationofpureconstituentsamples.Theamountofeachofthesesaltsneededtoobtain200mLofa5gL1solutionwascomputedandmeasuredandthereafterplacedinameasure-mentbottle.Approximately170mLdistilledwaterwasaddedtothebottleanditwasplacedonaheatedmagneticstirrer.Afterthesaltwasdeterminedtobecompletelydissolvedaccordingtoavisualassessmentadditionalwaterwasaddeduntilatotalvolumeof200mLwasobtained.Thesesolu-tionswerethereafterfurtherdilutedtoobtain50mLsamplesfrom1gL1to5gL1atastepsizeof1gL1.Solutionsfrom0.5gL1to4.5gL1atastepsizeof0.5gL1wereobtainedbydiluting10mLofthemastersolutionswithdistilledwaterintoatotalvolumeof200mLandthereafterfurtherdilutingthissolutioninto50mLsampleswithdistilledwatertothedescribedconcentrations.ThepureconstituentsampleswerethenplacedinarefrigeratorforlaterNIRspectroscopicmeasurements. 13J.DahlbackaT.LillhongaandM.DringJ.NearInfraredSpectrosc.2112222013ProcesssamplesTheprocesssampleswerecollectedfromfourbatchfermen-tations.Thesewerecarriedoutintwocustom-builtcontinu-ouslystirred38Llaboratoryscalereactorswithaworkingvolumeof27L.ThesetermophilicanaerobicdigestionswerestartedwithacultureobtainedfromanearbymunicipalwastetreatmentplantAb.StormossenOyandtheinoculumvolumewas3L.Thesubstratewasa4dryweightmixtureofpigmanurewastefromindustrialtreatmentofrawshandgreenhouseplantwasteatadryweightratioof2316and16.Duringthesefourdigestionsatotalof33sampleswereremovedfromthereactors.Duringasampleremovalthestirrerwashaltedforafewminutesandapproximately900mLofsamplewereremovedthroughthetopofthereactor.Thissamplewasthensplitinto16sub-samples14forthespectroscopicmeasurementsand2forthereferencemeas-urementsatavolumeof50mL.ThereafterthesampleswereplacedinafreezerforlaterNIRspectroscopicandreferencemeasurements.SpikingdesignThespikingoftheprocesssampleswascarriedoutasathree-factorcentralcompositedesignwithaxialpointsatthecubewallsandnocentrepoints.Thelengthofthecubewallwas2incodedspaceequivalentto5gL1inconcentrationspacewhenaddedtoasamplewithzeroconstituentconcentration.Thedesignitselfisfullyorthogonalbutbecomessomewhatskewedwhenimplemented.Thisoccursbecausetheprocesssamplesalwayscontainconstituentsofsomeconcentrationandthisinitialconcentrationwillbecomedilutedwhenspikingisperformed.Table1showsthedesignrepresentedascodedvaluesspikesizeinconcentrationspaceandtheamountofmastersolutionsneededtoperformthespikeona50mLprocesssample.Theprocesssampleasextractedfromthereactorrepresentsthecoordinate111inthecodeddesign.Intotalthe33processsampleseachonesplitinto14sub-samplesforNIRspectroscopicmeasurementsthenamountedto462samplesforNIRanalyses.ReferencemeasurementsTheammoniumconcentrationwasmeasuredusingowinjectionanalysisandphotometricaldetectionFIAstar5000AnalyserFossTecatorDenmarkbythegasperme-ablemembranemethodinaccordancewiththeENISO117322005procedure.17Thestandarderrorofthemethodwasdeterminedbyanalysingallsamplestwice.Thevolatilefattyacidconcentrationsweremeasuredusingagaschro-matographmassspectrometerShimadzuQP-2010GCMSShimadzuScientificInstruments7102RiverwoodDriveColumbiaMD21046USAinaccordancewiththemethod-ologydescribedbyYangandChoong.18TheTVFAconcentra-tionwascomputedastheionconcentrationsumofaceticacidpropionicacidisobutyricacidbutyricacid4-methylvalericacidvalericacid3-methylvalericacidandhexanoicacid.Thestandarderrorofthemethodwasdeterminedbyanalysingallsamplestwice.Theconstituentconcentra-tionsofthespikedsampleswerecomputedaccordingtotheamountofeachconstituentaddedineachspikeandthevolumeincreaseofthesamplethatthespikeresultedinaddedtothevolumeandtheresultsfromthereferencemeasurementsoftheprocesssamples.HardwareandsoftwareThemeasurementswerecarriedoutwithaportableHandySpecFielddiodearrayinstrumenttec5AGinderAu2761440OberurselGermany.TheinstrumentwasequippedwithanMMS1monolithicminiaturespectrometerdetectorforthelowerwavelengthsandaPGS2.2planegratingspectrom-eterdetectorforwavelengthsabove1000nmcomprising256sensorsintheregionfrom305nmto2200nm.TheAgroSpecsoftwaretec5AGinderAu2761440OberurselGermanywasusedastheinterfaceinthecollectionofthespectra.Anin-housebuilt5mmowthroughtransmissioncellwaspluggedinoneendandusedasacuvette.Picture1showstheequipmentused.Thereferencespectrumwascollectedondistilledwaterandtheintegrationtimewassetto7.5msforthePGSdetectorand4.5msfortheMMSdetector.Theseinte-grationtimesweredeterminedtobethemaximalallowableintegrationtimeswithwaterinthecell.Allspectraconsistedof32averagedscans.Thepartialleast-squaresPLSmodelswerecalculatedusingthePLS_Toolboxv.6.5.1EigenvectorResearchInc.3905WestEaglerockDriveWenatcheeWAPicture1.Instrumentationusedforthespectroscopicmeas-urementsshowingathewholesetupbthetransmissionheadandctheowthroughcelllledwithsample. 14OrthogonalSampleSpikingBasedCalibrationsforQuantitativeLiquidPhaseMeasurements98801USAtogetherwithMATLABR2011btheMathWorksABKistaSweden.SamplehandlingAllofthesamplesaccountedforinthisstudywereeitherstoredinarefrigeratororinthefreezerpriortoreferenceorspectroscopicmeasurements.Inthecaseofthepureconstituentsamplesthesampleswereremovedfromtherefrigeratorthedaybeforethespectroscopicmeasurementinordertoreachroomtemperature.Inthecaseofsamplesusedforreferencemeasurementsthesampleswerekeptinafrozenstateuntilitwastimetoperformthemeasurement.InthecaseoftheprocesssamplesandspikedsamplesthesamplesweremeltedinwaterbeforetheadditionofthespikesaccordingtoTable1wasperformedandthenthesampleswereallowedtoreachroomtemperatureovernight.Intherstseriesofspectroscopicmeasurementstheprocesssamplesandspikedsampleswereshakenandthereafterlledintothecuvette.Aftermeasurementthesampleswereagainplacedinthefreezer.Sinceitwaslaterconcludedthattheparticulatematterinthesamplesseverelyreducedtherepeatabilityofthemeasurementasecondseriesofspectroscopicmeas-urementswasperformed.Inthesecondseriesofspectro-scopicmeasurementsduringwhichallofthepureconstituentsampleswerealsomeasuredagaintheprocesssamplesandspikedsampleswerecentrifugedbeforethemeasurement.Thepureconstituentsamplesandthe462samplesoriginatingfromtheprocesssampleswereremovedfromtherefrigeratorandfreezerthedaybeforethemeasurementandallowedtoreachroomtemperatureovernight.Theprocesssampleswerethencentrifugedin2mLEppendorftubesat14000gfor5minandthecuvettewaslledwiththesupernatant.Threespectrawerecollectedfromeverysamplemeasuredinthisstudy.DatasetsAtotalof1476spectrarepresenting492sampleswerecollectedinthisstudy.Ninetyofthesespectrawerecollectedfromthe30pureconstituentsolutions.Thesespectraandcorrespondingconstituentdatawillhereafterbereferredtoasthepureconstituentdataset.Theremaining1386spectracamefrom33processsamplesand429spikedsamples.Basedontheobjectivetoobtainavalidationdatasetevenlydistrib-utedovertheconcentrationintervalofeachconstituentthe27spectraandthecorrespondingconstituentdatafromnineprocesssampleswereselectedasthevalidationdataset.Thisdatasetwillhereafterbereferredtoastheprocesssamplevalidationdataset.Thisdatasetwasfurtherextendedwiththe351spectraandthecorrespondingconstituentdatafromthe117spikedsamplesoriginatingfromthenineselectedprocesssamples.Thisdatasetisreferredtoasthespikedvalidationdataset.Theremaining1008spectraandthecorrespondingconstituentdatawillbecalledthespikedregressiondataset.Byextractingonlythe72spectraandtheconstituentdatarepresentingthe24processsamplesfromthisdatasetasetthatwillbecalledtheprocesssampleregressionsetwasobtained.ModelaccuracyTheaccuracyofthemodelsisdescribedbytherootmeansquareerrorofcalibrationRMSECtherootmeansquareerrorofcross-validationRMSECVtherootmeansquareerrorofthepredictionRMSEPtheratioofpredictiontostandarddeviationRPDtheratioerrorrangeRERandthecoefcientofdeterminationvaluesincrossvalidationR2ormodelvali-dationr2.ThedenitionoftherstthreecanbefoundinNsetal.19andthefollowingtwoinFearn.20Thecoefcientofdeterminationisthecorrelationsquared.ThersttwocanbesaidtoreecthowwellthemodelstthecalibrationdataTable1.Spikingdesigndescribedbycodedvaluesspikesizeinconcentrationspaceandspikesizefromthemastersolutionsinvolumespace.CodedcoordinatesSpikesizegL1SpikevolumemLAmmoniumAcetatePropionateAmmoniumAcetatePropionateAmmoniumAcetatePropionate1110.00.00.00.000.000.001110.00.05.00.000.002.631110.05.00.00.002.630.001110.05.05.00.002.782.781115.00.00.02.630.000.001115.00.05.02.780.002.781115.05.00.02.782.780.001115.05.05.02.942.942.940012.52.50.01.321.320.000012.52.55.01.391.392.780102.50.02.51.320.001.320102.55.02.51.392.781.391000.02.52.50.001.321.321005.02.52.52.781.391.39 15J.DahlbackaT.LillhongaandM.DringJ.NearInfraredSpectrosc.2112222013whereasthefollowingfourareusedtodescribetheaccuracyofthevalidationdatawhenavailable.ResultsanddiscussionSpectralqualityAsstatedearlierthefirstNIRmeasurementserieswascarriedoutonunpreparedprocessandspikedsamples.Themeasurementsconrmedtheassumptionthatthelowopacityofthesamplesandtheparticulatemattermakesthemeas-urementchallenging.Theconclusionwasthatitisnotadvis-abletoperformtransmittancemeasurementsonthistypeofsampleswiththismeasurementsetup.Thiswaspartlybasedonthelowqualityofsomeofthespectrabutthemainconcernwasthereproducibility.Duetothefairlyrapidsedimentationofthesolidphasethespectralfeaturesweresignicantlyaffectedbythetimethesamplehadbeenleftundisturbedbeforellingthecuvetteandthetimethesamplewaskeptinthecuvette.Thereforeallofthesamplesweremeasuredagainandthistimetheprocesssamplesandspikedsampleswerecentrifugedandmostofthesolidphasetherebyremovedbeforethemeasurement.Figure1showsthedetectorsignalandthecorrespondingabsorbancespectrumforoneprocesssamplefromthesecondmeasurementseries.AscanbeseeninFigure1thedetectorsignalforwavelengthsabove1400nmisrelativelylowalthoughthesamplehasbeencentrifuged.TheimpactofthespikingprocedureonthespectrallevelisillustratedinFigure2wherethelowestandthehighestabsorbancevaluesareshownfortheprocesssampleandspikedregressiondatasetsatdifferentwavelengths.Totheextentconclusionscanbedrawnbyvisualinspectionofthespectralinformationitisperhapsappropriatetosuggestthatthespikingproceduresomewhatwidenedtheabsorbancerangeinthemodelregressiondata.Howeveritisstillofaverysimilarmagnitudeincomparisontothatoftheprocesssampleregressiondataset.ConstituentintercorrelationOneobjectiveintheimplementationoftheorthogonalspikingschemewastoremoveconstituentintercorrelationfromthecalibrationdata.Table2containsthecorrelationmatrixforammoniumacetatepropionateandTVFAcomputedontheprocessandspikedsamples.Ascanbeseentheinter-correlationofthethreeconstituentsinthespikingschemewasessentiallyremovedbyapplyingthespikingprocedure.Howeveritcanalsobepointedoutthattheconstituentinter-correlationpriortothespikingofthesampleswaslowerthanwhatwasexpectedbasedontheliteraturevaluespreviouslyaccountedfor.Apossibleexplanationforthisisthattwoofthefermentationswereperformedwitha5gL1additionofvermiculitewhichhasbeenshowntoaffecttheammoniumconcentrationinanaerobicdigestionofsludge21andwhichalsoappearstoreducetheVFAconcentrationsatleastinFigure1.Detectorsignalblacklineandcorrespondingabsorbancespectrumgreylinecollectedfromthesupernatantofacentrifugedprocesssample.Figure2.Highestandlowestabsorbancevaluesintheprocesssampleblacklineandspikedgreylineregressiondatasets.Table2.Correlationmatrixforthestudiedconstituentsintheprocesssamplesbeforeaftertheimplementationofthedesignedspikingscheme.AmmoniumAcetatePropionateTotalVFAAmmonium1.001.00Acetate0.560.021.001.00Propionate0.570.010.490.021.001.00TVFA0.500.020.880.720.590.611.001.00 16OrthogonalSampleSpikingBasedCalibrationsforQuantitativeLiquidPhaseMeasurementsthecompostingofshwaste.22Onthewholetheobjectiveofremovingtheconstituentintercorrelationfromthesamplescanbesaidtohavebeenverysuccessfullycarriedoutfortheconstituentsinthespikingscheme.InitialinvestigationsTheobjectivetoevaluatetheimpactofacalibrationdatagenerationstrategyontheaccuracyofPLSmodelsisperhapsnotaverystraightforwardtaskduetothefactthatnumerousalternativesarealreadyathandinmostsoftwarewhenthemodelsareregressed.Thesealternativesincludeforexamplewavelengthselectionpre-processingmethodsandcombinationsofthesethenumberofPLScomponentstobeusedandremovalofoutliersinordertoimprovethemodelperformance.Initiallyitwasconsideredthattheimpactofthespikingprocedureonthemeasurementaccuracyshouldbeevaluatedbykeepingthemodelsettingsthesameforallmodels.Howeveritwasconcludedthatthisapproachwouldbelessthanreliable.Figure3showstheRMSECVandRMSEPvaluesasafunctionofthenumberofPLScomponentsusedfortwoacetatemodelsoneregressedontheprocesssamplesandoneregressedonthespikedsamples.AscanbeseenthenumberofPLScomponentsthatareselectedhaveadramaticimpactonthemodelaccuracy.ForinstanceifacomparisonwasbasedontheresultswithninePLScomponentsinbothmodelstheimpressionwouldbethatthemodelregressedonspikedsamplesisroughlytwiceasaccurateasthemodelregressedontheprocesssamples.ItwasthereforedecidedthatthecomparisonbetweenmodelsregressedonprocesssamplesandmodelsregressedonspikedsamplesshouldbemadebycomparingthebestmodelsfounddenedprimarilybyalowRMSEPvalue.Inordertoobtainadditionalinformationabouttheconstit-uentsstudiedandanindirectbench-markforthepredic-tionaccuracycalibrationmodelswereregressedonthepureconstituentspectraforallfourconstituents.Intheevaluationofthesemodelsnoexternalvalidationdatasetwasused.ConsequentlynoRMSEPvaluecouldbeobtained.InTables36thecoefcientofdeterminationbiasandslopevaluesthereforerefertotheregressiondataandtheRPDandRERvaluestotheregressiondataandtheRMSECVvaluesinthePureonPurecolumns.TheseexceptionsweremadeonlyforthemodelsregressedonpureconstituentspectraandforalltheothermodelsaccountedforthevariablesmentionedintheprevioussentenceexceptRMSECVrefertothevalida-tiondata.TheRMSECVvaluesinTables36showthatthehighestaccuracywasobtainedforammoniumandalsothattheaccuracyforTVFAwashigherthanforbothacetateandpropionate.Thiswasseenasanindicationofsignicantsimi-laritiesinspectralfeaturesbetweenacetateandpropionate.Inordertofurtherevaluatethisthespectrarepresentingzeroconstituentconcentrationwereexcludedfromtheregres-siondata.WiththesamemodelsettingsasdocumentedinTables36theRMSECVvaluesobtainedforammoniumacetatepropionateandTVFAwerethen215145withtwoPLScomponents93mgL1101mgL1and106mgL1respec-tively.Consequentlyasignicantportionofthemeasurementerrorforacetateandpropionatewhenregressedonallthepureconstituentspectracamefrominaccuratepredictionsonthenon-constituenti.e.zeroconcentrationspectra.Furthermoreitwasconcludedthatthespectralfeaturesofacetateandpropionatedonotreducethepredictionaccuracyforammoniuminthisdataset.Althoughitshouldbeacknowl-edgedthatthepureconstituentsamplesalsoincludeacertainlevelofconcentrationdiscrepancytheresultsfromthesemodelregressionsalsosuggestedthattheaccuracyfortheconstituentsandtheconcentrationrangewiththehardwareandmethodologyusedisunlikelytobebetterthan100mgL1.EvaluationofthespikingprocedureWhereasthemodelregressionsonpureconstituentspectragavesomevaluableinformationregardingobtainableaccu-racyandspectralsimilaritytheresultsdonotrevealanythingregardingtheusefulnessofthespikingmethodology.ThiswasinsteadevaluatedbyregressingmodelsontheprocesssampleregressiondataandbyvalidatingthemontheprocesssamplevalidationdatacolumnProc.onProc.inTables36byregressingmodelsonthespikedsampleregressiondataandbyvalidatingthemontheprocesssamplevalida-tiondatacolumnSpikeonProcinTables36.Theresultsfromthesecalibrationswereregardedasresultsfromtheactualapplicationasanoff-linemeasurement.Theaccu-racyshouldinthissensebecomparabletoliteraturevalueswhichisdonebyincludingreportedaccuraciesforNIRandbioprocessesingeneral23combinedwithresultsfromADNIRapplications2426inTables36.Inordertoevaluatehowcloselythemodelswererelatedtotheactualspectralfeaturesofthestudiedconstituentsthemodelsdescribedabovewerealsovalidatedagainstthepureconstituentspectra.ThisisdocumentedinTables36intheProc.onPureandSpikeonPurecolumns.ForeachconstituentmodelsregressedonspikedsampleregressiondatawerealsovalidatedagainstFigure3.Rootmeansquareerrorofcross-validationRMSECVgreylineandrootmeansquareerrorofpredictionRMSEPblacklineforacetatepartialleast-squaresPLScalibrationmodelsregressedonprocesssamplesoandspikedsamplesx. 17J.DahlbackaT.LillhongaandM.DringJ.NearInfraredSpectrosc.2112222013spikedvalidationdata.TheresultsfromthesearefoundinTables36intheSpikeonSpikecolumns.Thesemodelswereutilisedtoevaluatetheimpactoftheconstituentconcen-trationspanonthepredictionaccuracy.MeasurementsontheprocesssamplevalidationdatasetTheinitialcalibrationattemptsshowedthattheRMSEPvaluesforammoniumacetateandpropionatewerehighlydependentontheconstituentconcentrationrangeusedinthecalibrations.Thereforeonlyaconcentrationrangecomparabletotherangeintheprocesssampleswasusedinthecomparisonbetweenmodelsregressedonprocesssamplesandspikedsamples.Effectivelythismeansthatallsamplesrepresentingafullspikeoftheconstituentwereremovedfromtheregressiondatainadditiontosomehalfspikesappliedtosampleswithhighconstituentconcen-trationsalreadybeforespiking.HencethecalibrationdataTable3.PLScalibrationmodelsforammonium.ReferencemethodaccuracymgL122AccuracygiveninReference23convertedwhenneededtomgL1observations8Min4Max198Average78Median26PureonPureProc.onProc.SpikeonProc.SpikeonSpikeProc.onPureSpikeonPurespectraincalibration9072360100872360spectrainvalidation27273789090Preprocessing1der4521Autoscale1der3721Meancentre1der4521Meancentre1der4521Autoscale1der3721Meancentre1der4521AutoscalePLScomponents9991599RMSECmgL1988717073287170RMSECVmgL1155150186784150186RMSEPmgL117612785921711297r20.990.890.930.840.120.72BiasmgL10.288.148.8126.91572.227.5Slope0.9960.9590.8570.8560.0310.164RPD10.02.63.62.50.71.2RER32.39.112.77.62.33.9Table4.PLScalibrationmodelsforacetate.ReferencemethodaccuracymgL1161AccuracygiveninReferencess23and24convertedwhenneededtomgL1observations7Min280Max1476Average697Median600PureonPureProc.onProc.SpikeonProc.SpikeonSpikeProc.onPureSpikeonPurespectraincalibration9066178100866178spectrainvalidation27273789090Preprocessing1der4321Autoscale1der7321Autoscale1der6321Autoscale1der4321Autoscale1der7321Autoscale1der6321AutoscalePLScomponents11291529RMSECmgL1107385189450385189RMSECVmgL1265421214512421214RMSEPmgL133426053718412967r20.970.860.950.950.260.62BiasmgL1244410921912752798Slope0.9950.8811.1070.9220.2210.482RPD5.82.73.44.20.80.5RER18.97.19.113.52.71.7 18OrthogonalSampleSpikingBasedCalibrationsforQuantitativeLiquidPhaseMeasurementsforthespikedsamplesdidnotrepresentthefulldesignanymorebutneverthelessasetoforthogonalspikes.ForammoniumtheRMSEPforthemodelregressedonspikedsampleswas127mgL1whichwasalmost30lowerthanforthemodelregressedonprocesssamples.Incompar-isonwiththereferencemethodaccuracythiscanperhapsbeconsideredalargeerror.HoweverwhencomparedwiththeRMSECVobtainedforthepureconstituentsamplesitseemsunlikelythataconsiderabllyhigheraccuracycouldbeobtainedwiththeset-upused.ForacetatetheRMSEPforthemodelregressedonspikedsampleswas260mgL1whichwas20lowerthanforthemodelregressedonprocesssamples.Thiswasstill60higherthanthereferencemethoderrorbutalmostexactlythesameastheRMSECVobtainedonthepureconstituentsamples.ConsequentlythiswasconsideredasatisfactoryTable5.PLScalibrationmodelsforpropionate.ReferencemethodaccuracymgL1208AccuracygiveninReferences23and24convertedwhenneededtomgL1observations4Min206Max1364Average578Median370PureonPureProc.onProc.SpikeonProc.SpikeonSpikeProc.onPureSpikeonPurespectraincalibration9069345100869345spectrainvalidation21213509090Preprocessing1der3121Autoscale1der5521Meancentre1der4921Autoscale1der3321Autoscale1der5521Meancentre1der4921AutoscalePLScomponents10681568RMSECmgL1124182150498182150RMSECVmgL1249248157533248157RMSEPmgL125820362616161604R20.970.610.760.920.060.74BiasmgL110122393584831Slope0.9930.6820.7190.9820.0270.111RPD6.21.62.13.51.01.0RER20.14.45.69.63.13.1Table6.PLScalibrationmodelsforTVFA.ReferencemethodaccuracymgL1521AccuracygiveninRefs23and24convertedwhenneededtomgL1observations2Min1594Max2095Average1845Median1845PureonpureProc.onproc.Spikeonproc.SpikeonspikeProc.onpureSpikeonpurespectraincalibration9072378100272378spectrainvalidation24243399090Spectralrangenm90521051005210513051955130517551005210513051955Preprocessing1der5321Autoscale1der5321Meancentre1der5321Meancentre1der5321Autoscale1der5321Meancentre1der5321MeancentrePLScomponents91516101516RMSECmgL181373509529373509RMSECVmgL117110425825411042582RMSEPmgL18587047092316968r20.990.880.940.960.740.94BiasmgL115262882118611Slope0.9980.8690.8610.9390.8811.274RPD10.32.93.65.20.81.8RER29.37.59.122.22.25.2 19J.DahlbackaT.LillhongaandM.DringJ.NearInfraredSpectrosc.2112222013result.ForboththeacetatemodelstheRMSEPvaluescouldbesomewhatreducedbystepwiseremovingthecalibrationdatawiththehighestresiduals.ForpropionatetheRMSEPforthemodelregressedonspikedsampleswas203mgL1whichwas20lowerthanforthemodelregressedonprocesssamplesaswell.Aswasthecaseforacetatetheaccuracyofbothmodelscouldbeslightlyincreasedbyremovingcali-brationdatawithhighresiduals.Inthiscasetheaccuracyachievedwasactuallythesameasforthereferencemethodandevenhigherthanforthepureconstituentsamples.Howeverinonesensethiswasthepoorestresultforallofthestudiedconstituents.TwovalidationsampleswereremovedasoutlierswhichdecreasedtheconstituentconcentrationrangeaswellasthenumberofvalidationspectraandasanindirectresulttheRERandr2valuesforthemodelregressedonspikedsampleswereonly5.6and0.76andevenconsid-erablylowerforthemodelregressedonprocesssamples.Additionalvalidationsampleswouldthereforehavebeenveryvaluable.Howeverthisshortcomingisatleastpartlyover-comewiththeresultsfromthespikedvalidationdatasetwhichwillbeaccountedforlateron.ForTVFAtheRMSEPforthemodelregressedonspikedsampleswas704mgL1whichagainwasapproximately20lowerthanforthemodelregressedonprocesssamples.InthiscasetheRMSEPwasconsiderablylargerthantheRMSECVobtainedonthepureconstituentsamples.Howeveritisstillfullycomparablewiththereferencemethoderror.Hencethisresultwasdeemedasverysatisfactory.InFigure4thepredictionsonthevalidationdatafromtheseeightmodelsseecolumnsProc.OnProc.andSpikeonProc.inTables36areplottedagainstthereferencemeasurements.MeasurementsonthespikedvalidationandpureconstituentdatasetsSinceitwasconcludedthattheconstituentconcentrationrangesaffectedthemeasurementaccuracythisaspectwasevaluatedfurtherbyregressingcalibrationmodelsonallthespikedregressionsamplesandbyvalidatingthemagainstallthespikedvalidationsamples.TheseresultsarefoundinTables36intheSpikeonSpikecolumns.Generallythisexpansionoftheconstituentconcentrationrangeinthecali-brationandthevalidationdataincreasedtheRMSEPvaluebutatthesametimeitincreasedtheRERvalueaswell.OneexceptiontothisstatementwasobtainedforammoniumforwhichtheRERvaluedecreasedquitesignicantlybytheexpansionoftheconstituentconcentrationrange.Thereasonforthisisnotunderstood.AnotherexceptionwasobtainedforTVFAforwhichtheexpandedconstituentconcentrationrangedidnotaffecttheRMSEPvalue.AsaconsequencetheRERat22.2wasconsiderablyhigherthanforanyotherprocessorspikedsamplevalidationdatasetsandalmostashighasforthepureconstituentcalibration.ApossibleexplanationforthisisthatthepredictionaccuracyforTVFAwasmainlylimitedbythereferencemethodaccuracy.ThepredictionsforTVFAwiththismodelseecolumnSpikeonSpikeinTable6areplottedagainstthereferencevaluesinFigure5.OnFigure4.PLSmodelpredictionsvsreferencemeasurementsontheprocesssamplevalidationsetforaammoniumbacetatecpropionateanddtotalvolatilefattyacidsTVFAwithmodelsregressedonoprocesssamplesandspikedsamples. 20OrthogonalSampleSpikingBasedCalibrationsforQuantitativeLiquidPhaseMeasurementsthewholewiththeexceptionofTVFAanRERaround10wasobtainedforallmodels.SincethiswasgenerallyalsothecaseforthemodelsregressedonspikedsamplesandvalidatedagainstprocesssamplestheRMSEPvaluecanbesaidtobesomewhatproportionaltotheconstituentconcentrationrangestudied.ForpropionatethespikedcalibrationandvalidationdatasetsalsogaveamuchhighercorrelationandRERvaluethanforthemodelvalidatedagainstprocesssamples.Basedonthebehaviouroftheothercalibrationmodelsstudieditcanthereforebesuggestedthatthesomewhatdissatisfactoryresultsfromthevalidationontheprocesssampleswereduetoinsufcientandinaccuratevalidationdata.Inordertoevaluatehowcloselythepredictionsmadebythemodelswererelatedtothespectralfeaturesoftheactualconstituentvalidationswereperformedonthepureconstituentsamples.Thisrelationshouldbeofimportanceifpredictionisperformedonasamplewithamatrixunlikethematricesofprocesssamplecalibrationdata.Thevalidationsonthepureconstituentsamplesweremadewiththesamemodelsthatwerevalidatedagainsttheprocesssamples.Inthesevalidationstheamplitudeofthebiaswasgenerallysolargethatthemainparameterofinterestwasconsideredtobethecoefcientofdetermination.AscanbeseenfromcolumnsProc.onPureinTables36ther2isessentiallyzeroforallotherconstituentsthanTVFAforthemodelsregressedonprocesssamples.Ontheotherhandther2isfairlyhighforallofthemodelsregressedonspikedsamplesshownincolumnsSpikeonPureinTables36.Thusitwasconcludedthatthemodelsregressedonspikedsamplesmadepredictionsmorecloselyrelatedtotheactualspectralfeaturesoftheconstituent.Howevertoshowthetruebenetsofthispropertywouldrequiresampleswithmatrixcompositionsthatwerenotreadilyavailable.FinalremarksAsanalevaluationoftheresultsobtainedwiththespikingprocedureacomparisontovaluesfoundintheliteraturewasmade.Hereitshouldbepointedoutthatinthelitera-tureaccuracyforammoniumcontainsvaluesgivenbothasammoniumandammoniatheaccuracyforacetatecontainsvaluesgivenasacetateandaceticacidandtheaccuracyforpropionatecontainsvaluesgivenaspropionateandpropionicacidaswell.Forammoniumtheaccuracyobtainedinthisstudywasnotparticularlyhighincomparisonwiththelitera-turevalues.Howeveritseemsthattheconcentrationrangehasadramaticimpactonthepredictionaccuracyforthisconstituentwhichmightbepartoftheexplanationforthisresult.Foracetateandpropionatetheaccuracyobtainedontheprocesssamplevalidationsetcanbedescribedasexcel-lentincomparisonwiththeliteraturevaluesfoundandveryreasonableonthespikedsamplevalidationdataaswell.InthecaseofTVFAtheobtainedaccuracywasalsoexcellentincomparisonwithpreviouslyreportedvalues.Howeveritisacknowledgedthatthenumberoffactorsaffectingtheaccuracyindifferentapplicationswithdifferentinstrumentsanddifferentexperimentalsetupsissogreatthatthistypeofcomparisonperhapsmainlyrevealswhethertheobtainedaccuracyisreasonableornot.Onepotentiallyimportantparameterthatwasnotevaluatedanyfurtherinthisworkisthespikesize.Theincrementsusedhereextendedthecalibrationintervalsintermsofconcentra-tionspanssignicantly.Thevalidationofthemodelsshowedthatahigheraccuracywasobtainedwhenthecalibrationdatawithsignicantlyhigherconcentrationthanintheprocesssampleswasomitted.Itisthereforesuggestedthatitcouldhavebeenbenecialtospikethesampleswithincrementsthatrepresentroughly10oftheconstituentconcentra-tionspaninthestudiedapplicationandtherebyachieveagoodcoverageandlimitedextrapolationoftheconcentrationintervalofinterest.Neverthelesstheresultswereencour-agingenoughtosuggestthatthemethodologyusedshouldbeofinterestforapplicationswheresevereconstituentinter-correlationispresentinthecalibrationsampleswherethecostsofperformingreferencemeasurementareveryhighandorwhereperformingreferencemeasurementsandcollectionofreferencesamplesisverytime-consuming.Furthermoretheaccuracyachievedforallofthefourconstituentsstudiedisperceivedashighenoughtomakethisequipmentinter-estingforoperatorsofADunits.Althoughthemeasurementswereperformedoff-linethemethodologyusedenablesthedeterminationoffourconstituentconcentrationssimulta-neouslywithinroughly5minfromsampleremovalfromthereactor.ConclusionsTheusefulnessofperformingspikingaccordingtoacentralcompositedesignforcalibrationpurposesonsamplesfromananaerobicdigestionprocesswasinvestigated.Itwasshownthattheprocedurecompletelyremovedtheintercor-relationbetweentheconstituentsincorporatedinthespikingdesign.Thisapproachtherebyalsochangestheconstit-uentandspectralcorrelationinspikedsamplesremovingFigure5.PLSmodelpredictionsforTVFAvsreferencemeas-urementforthemodelregressedonspikedsamplesandvalidatedagainstspikedsamples. 21J.DahlbackaT.LillhongaandM.DringJ.NearInfraredSpectrosc.2112222013potentiallyusefulcorrelation.Howevercalibrationmodelsregressedonthespikedsamplesconsistentlyperformedbetterthantheonesregressedonprocesssamples.OnaveragetheRMSEPwasreducedby24forammoniumacetateandpropionate.Validationsofthemodelsagainstpureconstituentspectraalsoshowedamuchhighercorre-lationforthemodelsregressedonspikedsamplesthanformodelsregressedonprocesssamples.Henceitcanbesaidthatthespikingprocedureresultedinmodelswhosepredictionsaremorecloselyrelatedtoactualspectralfeaturesoftheconstituents.Thisshouldbebeneficialincasetheunknownspectrarepresentamatrixthatdoesnotfallwithinthecalibrationdatamatrix.Thusthegeneralconclusionisthatthemethodologyshouldbeofinterestinapplicationswhereperformingthereferenceanalysesonthecalibrationdataareverycostlyintimeormoneyorother-wiseasamethodtoremoveundesirableintercorrelation.Fromameasurementapplicationperspectivetheaccuracyobtainedinthisstudyshouldbehighenoughtomaketheequipmentappealingtooperatorsofanaerobicdigestionplants.WithNIRspectroscopyitwaspossibletodeterminefourconstituentconcentrationswithasingleinstrumentatareasonableaccuracyandtherebyatareasonablecost.AcknowledgementsThisworkwasmadepossiblethroughtheFIELD-NIRceBio-BioandMarePurumprojectsfundedbytheEuropeanterritorialcooperationprogrammeBotnia-Atlantica.TheauthorsalsowanttothankProfessorPaulGeladiforscienticadvice.References1.P.WeilandBiogasproductioncurrentstateandper-spectivesAppl.Microbiol.Biotechnol.8548492010.doi10.1007s00253-009-2246-72.A.J.WardP.J.HobbsP.J.HollimanandD.L.JonesOptimisationoftheanaerobicdigestionofagriculturalresourcesBioresour.Technol.991779282008.doi10.1016j.biortech.2008.02.0443.D.WolfH.vonCansteinandC.SchrderOptimisationofbiogasproductionbyinfraredspectroscopy-basedprocesscontrolJ.Nat.GasSci.Eng.36252011.doi10.1016j.jngse.2011.07.0064.M.MadsenJ.B.Holm-NielsenandK.H.EsbensenMonitoringofanaerobicdigestionprocessesAreviewperspectiveRenew.Sustain.EnergyReviews1531412011.5.J.B.Holm-NielsenH.AndreeH.LindorferandK.H.EsbensenTransexiveembeddednearinfraredmoni-toringforkeyprocessintermediatesinanaerobicdiges-tionbiogasproductionJ.NearInfraredSpectrosc.151232007.doi10.1255jnirs.7196.J.R.DeanMethodsforEnvironmentalTraceAnalysis.JohnWileySonsChichesterEngland2003.7.L.ErikssonE.JohanssonN.Kettaneh-WoldC.WikstrmandS.WoldDesignofExperimentsPrinciplesandApplications.UmetricsABUme2008.8.B.FinnL.M.HarveyandB.McNeilNear-infraredspectroscopicmonitoringofbiomassglucoseethanolandproteincontentinahighcelldensitybakersyeastfed-batchbioprocessYeast2375072006.doi10.1002yea.13719.M.R.RileyM.RhielX.ZhouM.A.ArnoldandD.W.MurhammerSimultaneousmeasurementofglucoseandglutamineininsectcellculturemediabynearinfra-redspectroscopyBiotechnol.Bioeng.551111997.doi10.1002SICI1097-0290199707055513.0.CO2-10.M.R.RileyM.A.ArnoldandD.W.MurhammerMatrixenhancedcalibrationprocedureformultivariatecalibra-tionmodelswithnearinfraredspectraAppl.Spectrosc.5213391998.doi10.1366000370298194267211.M.R.RileyM.A.ArnoldD.W.MurhammerE.L.WallsandN.DelaCruzAdaptivecalibrationschemeforquanticationofnutrientsandbyproductsininsectcellbioreactorsbynear-infraredspectroscopyBiotechnol.Prog.1435271998.doi10.1021bp980022d12.K.S.Y.YeungM.HoareN.F.ThornhillT.WilliamsandJ.D.VaghjianiRealtimemonitoringandcontrolofglucoseandlactateconcentrationsinamammaliancellperfusionreactorBiotechnol.Bioeng.6366841999.doi10.1002SICI1097-0290199906206363.0.CO2-Q13.P.RoychoudhuryB.McNeilandL.M.HarveySimultaneousdeterminationofglycerolandclavulanicacidinanantibioticbioprocessusingattenuatedtotalreectancemidinfraredspectroscopyAnal.Chim.Acta58522462007.doi10.1016j.aca.2006.12.05114.V.G.FrancoJ.C.PernV.E.MantovaniandH.C.GoicoecheaMonitoringsubstrateandproductsinabioprocesswithFTIRspectroscopycoupledtoarticialneuralnetworksenhancedwithagenetic-algorithm-basedmethodforwavelengthselectionTalanta68310052006.doi10.1016j.talanta.2005.07.00315.P.F.PindI.AngelidakiandB.K.AhringAnewVFAsen-sortechniqueforanaerobicreactorsystemsBiotechnol.Bioeng.821542003.doi10.1002bit.1053716.Y.ChenJ.J.ChengandK.S.CreamerInhibitionofanaerobicdigestionprocessareviewBioresour.Technol.991040442008.doi10.1016j.biortech.2007.01.05717.WaterQuality.DeterminationofAmmoniumNitrogen.MethodbyFlowAnalysisCFAandFIAandSpectrometricDetectionSFS-ENISO11732.FinnishStandardsAssociationSFSHelsinkiFinland2005.18.M.-H.YangandY.-M.ChoongArapidchromatographicmethodfordirectdeterminationofshort-chainC2C12volatileorganicacidsinfoodsFoodChem.751012001.doi10.1016S0308-81460100211-419.T.NsT.IsakssonT.FearnandT.DaviesAUserFriendlyGuidetoMultivariateCalibrationandClassication.NIRPublicationsChichesterUK2002. 22OrthogonalSampleSpikingBasedCalibrationsforQuantitativeLiquidPhaseMeasurements20.T.FearnAssessingcalibrationsSEPRPDRERandR2NIRnews136122002.doi10.1255nirn.68921.C.TadaY.YangT.HanaokaA.SonodaK.OoiandS.SawayamaEffectofnaturalzeoliteonmethaneproductionforanaerobicdigestionofammoniumrichorganicsludgeBioresour.Technol.9644592005.doi10.1016j.biortech.2004.05.02522.P.H.LiaoL.JonesA.K.LauS.WalkemeyerB.EganandN.HolbekCompostingofshwastesinafull-scalein-vesselsystemBioresour.Technol.591631997.doi10.1016S0960-85249600153-823.A.E.CerveraN.PetersenA.E.LantzA.LarsenandK.V.GernaeyApplicationofnear-infraredspectroscopyformonitoringandcontrolofcellcultureandfermentationBiotechnol.Prog.25615612009.24.C.LomborgJ.B.Holm-NielsenP.Oleskowicz-PopielandK.H.EsbensenNearinfraredandacousticchemo-metricsmonitoringofvolatilefattyacidsanddrymat-terduringco-digestionofmanureandmaizesilageBioresour.Technol.100517112009.doi10.1016j.biortech.2008.09.04325.M.HanssonA.NordbergI.SundhandB.MathisenEarlywarningofdisturbancesinalaboratory-scaleMSWbiogasprocessWaterSci.Technol.45102552002.26.M.HansonA.NordbergandB.MathisenOn-lineNIRmonitoringduringanaerobictreatmentofmunicipalsolidwasteWaterSci.Technol.48492003. JOURNALOFNEARINFRAREDSPECTROSCOPY23ISSN0967-0335IMPublicationsLLP2013doi10.1255jnirs.1033AllrightsreservedLocalcalibrationmodelscanbesaidtobefairlycommonlyusedwithintheeldofnearinfraredNIRspectroscopy.InthiscontexttherstexampleontheuseoflocalcalibrationtechniquesisprobablythatofDaviesetal.1Localalgorithmscanbebenecialbyreducingtheimpactofapotentiallynon-linearrelationshipbetweentheconstituentandthespectralinformationaswellasofsampleinhomogeneity.2Howeveralthoughthishasbeenthetopicofnumerouspublicationsthereappearstobenogenerallyaccepteddenitionofwhatalocalcalibrationtechniqueis.FromtheworkofFearnandDavies3theinterpretationcanbemadethatalocalmodelisamodelforwhichthecalibrationdatahasbeenselectedbasedonanestimateofclosenessfromalargerdatabaseandthatthepredictionisbasedontheinforma-tionfromtheseselectedspectra.Withthisinterpretationtheconstructionofalocalcalibrationmodelbecomesatwo-stepoperationinwhichtherstistoselectthelocalsubsetofcalibrationdataandthesecondistoextractapredictivemodelfromthisinformation.Itisinthesetwostepsthatthedifferencebetweenlocalcalibrationtech-QuantitativemeasurementsofanaerobicdigestionprocessparametersusingnearinfraredspectroscopyandlocalcalibrationmodelsJohnDahlbackaabandTomLillhongaaaNoviaUniversityofAppliedSciencesPOBox6FIN65201VaasaFinland.E-mailJohn.Dahlbackanovia.Theperformanceoflocalcalibrationmodelsforquantitativemeasurementsofammoniumandacetateonsamplesfromananaerobicdigestionprocesswasexamined.ThelocalcalibrationmethodsusedwerelocallyweightedregressionLWRandmulti-layerpartialleastsquaresML-PLSregression.TheresultsofthesetwomethodswerecomparedtoeachotherandtotheresultsfromtheglobalpartialleastsquaresPLSmodelregressionaswell.ForammoniumboththelocalmethodsperformedexcellentlyincomparisonwithglobalPLSmodels.Howevertheresultsfromthe150LWRmodelsregressedforammoniumalsoshowedthattheaccuracycanbehighlydependentonthedifferentcombinationalternativesformodelparametersettingsandpre-processingalternatives.ForthisreasonanumberofdistancemeasureswereevaluatedaslocalsubsetselectionmethodsinML-PLS.ThebenetsofanoptimisedlayerstructureandtheiterativeapproachinML-PLSwerealsoevaluatedforammonium.Thisshowedthatsomebenetscanbeobtainedbyoptimisingthelayerstructureatleastinthesensethatthenumberoflayerscanbereducedandthattherecanbeasignicantadvantageinusinganiterativeapproachintheselectionofthelocalsubsetofcalibrationdata.ThelocalcalibrationmethodswerealsoevaluatedforacetatebutinthiscasethebenetscomparedtoglobalPLScalibrationmodelswerefairlyinsignicantwithML-PLSandnoneatallwithLWR.KeywordsanaerobicdigestionbioprocessmetabolitesnearinfraredNIRlocallyweightedregressionLWRlocalmodelmulti-layerpartialleastsquaresML-PLSlocalcalibrationammoniumIntroductionJ.DahlbackaandT.LillhongaJ.NearInfraredSpectrosc.2123332013Received29June2012Revised17October2012Accepted30October2012Publication21January2013 24QuantitativeMeasurementsofADProcessParametersniquesusedinNIRspectroscopycanbefound.InCARNAC1thesimilaritybetweentheunknownsampleandtheavailablemodelregressiondataisevaluatedbymeansofcorrelationsquared.HoweverthecorrelationiscomputedonweightedFouriercoefcientsandnotontheNIRspectra.TheCARNACalgorithmsdonotincorporateanactualmodelregressionstep.Insteadthepredictioniscalculatedfromasimilarityweightedaverageoftheselectedlocalsubsetofsamples.Locallybiasedregression3isamethodthatusestwocriteriatoassessthesimilaritybetweentheunknownspectrumandthedatabase.Therstcriterionisthedistancetothepredic-tionmadebyaglobalpartialleastsquaresPLSmodelandthesecondisthatthescoreontherstorthogonalsignalcorrectionOSCfactorshouldbeinsideaspeciedintervalcomparedtotheunknownspectrum.ThereaftertheactualpredictionismadebyeitherbiasorskewandbiascorrectionofthepredictionmadebytheglobalPLSmodel.Thebiasandskewinturnaredeterminedfromtheperformanceoftheglobalmodelontheselectedsubsetofspectra.TheLOCAL4algorithmselectsthelocalsubsetofcalibrationdatabasedonthespectralcorrelationtotheunknownwhereafteraPLSmodelisregressedontheselectedspectra.ThemostrecentalgorithmthatwasfoundiscalledthelocalcentralalgorithmLCA5andinthisthedistanceinprincipalcompo-nentPCspaceisusedtodeterminesimilarity.Themethodusesnocalibrationstepinsteadthepredictionisbasedontheaveragemedianormodeoftheselectedsubset.Inadditiontothesemethodsthereisthewell-knownmethodoflocallyweightedregressionLWRandamethodologycurrentlybeingdevelopedcalledmulti-layerPLSML-PLS.Sincethesetwomethodsareevaluatedinthisworkamoredetaileddescriptionisgivenbelow.FurtherinformationoncalibrationmethodsusedinNIRspectroscopycanalsobefoundintheliterature.67TheoriginofLWRcanbefoundintheworkofCleveland8inwhichrobustnessofsmoothingthroughpolynomialttingbymeansofweightedleastsquareswasexamined.ClevelandandDevlin9expandedthemethodologyamongotherthingsasamethodoflinearmodellingofnonlinearregressionsurfaces.Nsetal.10introducedLWRtoanddevelopeditfortheeldofvibrationalspectroscopy.Thesuggestedconceptforpredictionofunknownsampleswastoselectaspeci-ednumberofpointsinthecalibrationsetthatwasclosesttothepredictionspectrumandbasetheregressionsurfaceonthesesamplesi.e.makeaweightedregressioncalibra-tionforeachnewpredictionsample.ClosenesswasinthiscasedefinedasMahalanobisdistanceinPCspace.Latersuggestionsforimprovementofthismethodincludemainlytheuseofnewdistancemeasuresandmodels.NsandIsaksson11suggestedthatthepredictionabilityofthePCsshouldbetakenintoaccountintheselectionofthesetofclosestcalibrationpoints.Wangetal.12exploredtheperfor-manceofincorporatingthedistancetothedependentvariableinthedistancemeasure.Inthismethodologythedistancetothedependentvariableiscanbecomputedaccordingtoaniterativeprocedurewherethepredictionmadebythemodelinthepreviousiterationdeterminesthedistancetothedependentvariableinthenextiterativeselectionoftheclosestcalibrationpoints.AastveitandMarum13suggestedthatprincipalcomponentregressionPCRcanbecarriedoutdirectlyontheselectedlocalsubsetratherthanontheglobalPCsandShenketal.4proposedthatPLSisusedforthelocalcalibrationmodel.ThesoftwareusedtocreateLWRmodelsinthispresentstudyusestheEuclideandistancetodeterminetheclosenesstotheconstituentordependentvariabletheauto-scaleddistanceinPCspacetodeterminetheclosenesstothespectrumandenablesmodelregressiononglobalPCslocalPCRandPLSregression.Theweightofthedistancetopredictionandthenumberofiterationsinselectingthelocalsubsetcanalsobespecied.Howeveritappearsthatalloftheselectedspectraareweightedequallyinthemodelregressionstep.DahlbackaandLillhonga14usedalocalmodelregressiontechniquecalledML-PLSforenhancedaccuracyinquanti-tativemoisturemeasurementsintimberwithNIRspectro-scopy.ThenamederivesfromtheinitialideathatthemodelconsistsofastructureofPLSmodelsorganisedonlayersinwhichthepredictiononthepreviouslayerservesasaselectionmethodforselectingwhichmodelwilldothepredictiononthenextlayer.Thespecicationforthelayerswasthatforeachlayerthenumberoflocalpointsusedformodelregressionshoulddecrease.Thusitisessentiallyamethodthatperformsastepwisereductionofthelocalsubsetofcalibrationsamples.Asitcanbeseenasindif-ferentwhethertheprocedureselectsanewmodelbasedonareducedregressiondatasetorwhetheritreducestheregressiondatasettobasethemodelonML-PLScanalsobeseenasapurelyiterativeproceduretoselectthelocalsubsetofcalibrationsamples.FurthermoretheselectioncriteriondoesnothavetobebasedonaPLSmodelpredic-tionwhichwasthecaseintheoriginalworkandwhenimplementedasaniterativeregressionmethodmultipleselectioncriteriacanbeusedsimultaneously.FinallythereisnothinginthemethodologythatdictatestheuseofaPLSmodel.ConsequentlythefeaturethatcouldbeperceivedasuniquewithML-PLSisthestepwiseoriterativereductionoftheregressiondatasetwhereasothermethodsseektondadatasetofdenedandoptimalsize.AsitwasassumedthatthisfeaturecouldhavesomebenefitscomparedtoselectingalocalsubsetofadenedsizeandthatitmightbeanovelmethodtoselectthelocalsubsettheconceptofML-PLSisappliedbroadenedandfurtherinvestigatedasaregressionmethodforquantitativemeasurementsonsamplesfromananaerobicdigestionADprocessinthispresentstudy.ThisisdonebytestingdifferentselectioncriteriaandcombinationsofthesebyevaluatingtheimpactofthestepwiseprocedureandbycomparingthendingstothecloselyrelatedLWRmethod.AsthisstudyfocusessolelyontheconceptoflocalmodelsotherpublicationscanberecommendedtoobtainknowledgeabouttheimpactandusefulnessofNIRspectroscopyforprocessmonitoringpurposesinAD.1521 J.DahlbackaandT.LillhongaJ.NearInfraredSpectrosc.212333201325MaterialsandmethodsSamplesTheprocesssampleswerecollectedfromfourbatchfermen-tationseachoneapproximatelyfourweekslong.Thesewerecarriedoutintwocustom-builtcontinuouslystirred38Llabo-ratoryscalereactorswithaworkingvolumeof27L.ThethermophilicADswerestartedwithacultureobtainedfromanearbymunicipalwastetreatmentplantandtheinoculumvolumewas3L.Thesubstratewasa4dryweightmixtureofpigmanurewastefromindustrialtreatmentofrawshandgreenhouseplantwasteatadryweightratioof2316and16.Duringthesefourdigestionsatotalof33samplesweretakenfromthereactorssplitinto14subsamplesforspikingandspectroscopicmeasurementsandthereafterplacedinafreezer.Thesesampleswerelaterspikedaccordingtoanorthogonalspikingschemewhichresultedin14sampleswithauniqueconcentrationcompositionforeachreactorsample.ThespikingofthesamplesisdescribedindetailinDahlbackaetal.22AsaresultofthespikingprocedurethetotalamountofsamplesthatweremeasuredwithNIRwas462.Beforethespectroscopicmeasurementthesampleswereremovedfromthefreezerandallowedtoreachroomtemperatureovernight.Thesampleswerethencentrifugedin2mLEppendorftubesat14000gfor5minandthespectroscopicmeasurementcarriedoutonthesupernatant.ReferencemeasurementsReferencemeasurementswereonlyconductedontheprocesssamplesandtheconcentrationsinthespikedsampleswerecomputedfromthespikesizeandtheconcentrationintheunspikedsample.Theammoniumconcentrationwasmeas-uredusingowinjectionanalysisandphotometricaldetec-tionFIAstar5000AnalyzerFossTecatorDenmarkbythegaspermeablemembranemethodinaccordancewiththeENISO117322005procedure.23Thestandarderrorofthemethodwasdeterminedbyanalysingallprocesssamplestwice.ThevolatilefattyacidsconcentrationsweremeasuredusingagaschromatographmassspectrometerShimadzunQP-2010GCMSShimadzuScientificInstruments7102RiverwoodDriveColumbiaMD21046USAinaccordancewiththemethodologydescribedbyYangandChoong.24Thestandarderrorofthemethodwasdeterminedbyanalysingallprocesssamplestwice.Theaccuracyforammoniumwas22mgL1andforacetate161mgL1.DatasetsThreespectrawerecollectedforeachofthe462samplesresultinginatotalof1386spectraavailableformodelregres-sionandvalidation.Fromtheoriginal33reactorsamplesbasedonanobjectivetocoverthefullconcentrationrangesevenlyninesampleswiththecorrespondingspikedsubsampleswereselectedasvalidationsamples.Thusthevalidationdatasetconsistedof378spectrafrom126reactororspikedsamples.Thesespectraandcorrespondingconcentrationsarereferredtoasthevalidationdatasetorsometimesforadditionalclari-cationthefullvalidationdataset.Inordertoreducethetimeneededtovalidatethemodelsasmallervalidationdatasetwasextractedfromthisfullvalidationdataset.Thisdatasetcomprised50spectraandcorrespondingconcentrationsbutwithnoreplicatesselectedfromthefullvalidationdataset.Thisdatasetisreferredtoasthereducedvalidationdataset.Theremainingdatai.e.whatwasnotincludedinthevalida-tiondatasetwasusedformodelregressions.Thisdatasetcontained1008spectrafrom336samplesoriginatingfrom24reactorsamplesandisreferredtoastheregressiondataset.LocalsubsetselectionmethodsForML-PLSlocalsubsetselectionmethodsbasedonabso-luteEuclideanandMahalanobisdistancesaswellascorrela-tionsandpredictionresidualswereevaluated.ThedenitionoftheEuclideanandMahalanobisdistancescanbefoundinNsetal..25TheEuclideandistancetopredictionwascomputedastheEuclideandistancebetweenthepredictionmadeonthepreviouslayerandtheconstituentreferencevaluesinthecalibrationdataset.TheEuclideandistanceinspectralspacewascomputedastheEuclideandistancebetweentheunknownspectrumandthespectrainthecalibrationdataset.Inordertoreducetheimpactofbaselinefeaturesthespectrawerepre-processedwitharstorderSavitzkyGolayderivative262ndorderpolynomial45pointspriortothecomputationofthisparameter.TheEuclidiandistanceinPLSspacewascomputedastheEuclideandistanceinPLSscorespaceforthesamenumberofcomponentsasinthePLSmodel.Theabsolutedistanceinspectralspacewascomputedonspectrapre-treatedthesamewayasfortheEuclideandistanceinspectralspace.TheabsolutedistanceinPLSspacewascomputedanalogouswiththeEuclideandistanceinPLSspaceaswasthecasewiththeMahalanobisdistanceinPLSspace.Thecorrelationinspectralspacewascomputeddirectlyontheabsorbancespectrai.e.withnopre-processing.ThecorrelationinPLSspacewascomputedforthesamenumberofPLSscoresasthenumberofPLScomponentsusedinthePLSmodel.Theregressionresidualwascomputedastheabsolutedeviationofpredictiontothereferencevaluewhenpredictingthecalibrationdataset.SpectroscopyThemeasurementswerecarriedoutwithaportableHandySpecFielddiodearrayinstrumenttec5AGinderAu2761440OberurselGermany.TheinstrumentwasequippedwithanMMS1monolithicminiaturespectrometerdetectorforthelowerwavelengthsandaPGS2.2planegratingspectrometerdetectorforwavelengthsabove1000nmcomprising256sensorsintheregion305nm2200nm.Howeverthespectrawerestoredwitha1nmspectraldataresolution.Thespectrawerecollectedintrans-missionmodethrougha5mmcuvette.Thereferencespectrumwascollectedonwaterandtheintegrationtimewassetto7.5msforthePGSdetectorand4.5msfortheMMSdetector.Allspectraconsistedof32averagedscans.ThePLSmodelswerecalculatedusingthePLSToolboxv.6.5.1EigenvectorResearchInc.3905WestEaglerockDriveWenatcheeWA98801USAtogetherwith 26QuantitativeMeasurementsofADProcessParametersMATLABR2011btheMathWorksABKistaSweden.ML-PLSregressionandvalidationwereimplementedasMATLABscriptscallingtheToolboxfunctions.SimilarlytheLWRmodelregres-sionsandvalidationswerealsocarriedoutthroughMATLABscriptsutilisingtheLWRfunctionsoftheToolbox.Forallmodelsaccountedforinthisstudytherstspectralpre-treatmentwasarstorderSavitzky-Golayderivativebasedon45pointswitha2ndorderpolynomial.Intheammoniumcalibrationsthespec-tralinterval1155nm1905nmwasusedinthemodelsandintheacetatecalibrationstheintervals905nm1405nmand1545nm1755nm.TheperformanceofthemodelsisassessedbasedontherootmeansquareerrorofpredictionRMSEP.25UnlessotherwisestatedallmodelsincludedsevencomponentsPLSorPC.WheneverthedistancetopredictionwasincludedasaselectioncriterioninLWRveiterationswereused.ResultsanddiscussionInitialresultswithML-PLSandammoniumTherstammoniumML-PLSmodelthatwasregressedwasbasedonauto-scaledspectralandconstituentdataandadesignwith99layersinwhich10regressionspectrawereremovedfromeachlayeraccordingtotheEuclideandistancetopredictioncriterion.WiththismodeltheRMSEPontherstlayeri.e.theglobalmodelwas873mgL1thelowestRMSEPof633mgL1wasobtainedatlayer67andonthelastlayertheRMSEPwas722mgL1.Thismaynotbeaparticularlyencouragingresultbutontheotherhandthescatterplotofthelastlayerindicatedthatmostofthemeasurementerrororiginatedfromafewverypoorlypredictedspectradatanotshown.Byperformingasimpleoutlierdetectionthroughreplicateanalysisstatingthatnosinglepredictioncanhavearelativeabsolutedeviationfromthemedianpredictionofthethreereplicatestothemedianpredictionofthethreereplicatesexceeding30theRMSEPofthelastlayerwasreducedto643mgL1.Byfurtherremovingthesevensamples21spectrathatgavethehighestcontributiontotheRMSEPtheRMSEPonthelastlayercouldbereducedto296mgL1.ThusforammoniumthereappearedtobeagreatpotentialforimprovementofthemeasurementaccuracywithML-PLScomparedtoaglobalPLSiftheseapparentoutlierswerenotpresentinthevalidationdata.Thedilemmawasthere-foretoevaluateifthesedatarepresentedtrueoutliersorifitwasmerelyaresultoftheregressionmethodused.IntheglobalmodelnothingextraordinarywasobservedintermsofHotellingsT227Qresidual27predictionresidualorscoresonthePLScomponents.AlsotheremovalofthesesevensampleshadnoeffectontheRMSEPoftheglobalmodel865mgL1.Localmodelswerealsoregressedaroundthesepointsbutthepredictionsbythesewereveryreasonable.Thebehaviourofthepredictionsfromlayertolayerwasalsoinves-tigatedinordertogetanunderstandingofwhythesepointsresultinpoorpredictionaccuracy.Itwasevaluatediftheabso-luteorrelativeabsolutechangeinthepredictionfromthersttothelastlayerwasextraordinary.Itwasnot.Itwasevaluatediftheabsoluteorrelativeabsolutechangeinthepredictioninthelastquarterofthedesignwasextraordinary.Itwasnot.Itwasevaluatediftheabsoluteslopethroughthelayersandforthelastquarteroflayerswasextraordinary.Itwasnot.Othermethodswhichwillnotbeaccountedforherewerealsoutilisedtoidentifythesepointsbuttheendresultwasthatnoclearindicationthatthesepointswereactuallyoutlierscouldbefound.Withanexpectationthatthisissuecouldbesolvedduringtheprogressofthisstudyitwasdecidedthatnoneoftheseapparentoutlierswouldbeincludedinthereducedvali-dationdatasetforammonium.Insteadthissetwasobtainedbyrstremovingtheapparentoutlierssecondbyremovingallreplicatespectraandthirdbyaimingatanevenlycoveredconstituentintervalinthevalidationdata.LWRandammoniumTheperformanceofLWRonthereducedvalidationsetcanbefoundintermsofRMSEPvaluesinTable1.Onthisvalida-tionsettheRMSEPoftheglobalPLSmodelwas856mgL1onmeancentredspectraand812mgL1onauto-scaledspectra.IncomparisonthelowestRMSEPobtainedwithLWRwas249mgL1forPCRanda0.75whereaistherelativeweightoftheEuclidiandistancetopredictionintheselec-tionofthesubsetofcalibrationspectra.OnamoregenerallevelitcanbesaidthatLWRperformedpoorlyonmeancentredspectraldataandwithanycombinationofparam-etersfora1.00.Consideringthatthefullregressionsetcontained1008spectraitcanalsobesaidthatthelowesterrorsareobtainedforarelativelyfewnumberofspectrainthelocalsubsetofcalibrationdata.ThehighestRMSEPwasalsoeighttimeshigherthanthelowestRMSEPobtainedwithLWRwhichindicatesthatathoroughandmethodologicalscreeningofregressionparametersisessentialinordertoevaluatethetruebenetsofLWR.Inthiscase150modelswereregressedwhichshouldprobablygiveanadequatecoverageofthepossibleregressionparametercombinations.OnthewholeLWRperformedexcellentlyonthisvalidationdatasetreducingtheRMSEPto29ofthatoftheglobalPLSmodelwiththesamespectralpre-treatment.SelectionmethodsinML-PLSforammoniumThedifferentdistancemeasuresforselectingthenextlocalsubsetofcalibrationspectrainML-PLSwerealsoevalu-atedonthereducedvalidationdataset.Inthesetestsa56layerdesignwasusedand18spectrawereremovedforeachlayeraccordingtothedistancemeasureused.TheresultsareshowninTable2intermsofRMSEPandRMSEPrelativetotheRMSEPontherstlayerinthedesignoneithermeancentredorauto-scaledspectraldata.Themin-valuesinTable2arethelowestRMSEPvaluesobtainedonanylayerinthedesigni.e.notnecessarilyonthelastlayer.AccordingtoTable2noimprovementinaccuracywasobtainedwhentheregressionspectrawereremovedbasedonpredictionresidual.Thiscrite-rionisperhapsnotconsistentwithanydenitionoflocalcali-brationtechniquesbutontheotherhandtheobtainedresultisavaluableindicatorthattheregressiondatasetdidnot J.DahlbackaandT.LillhongaJ.NearInfraredSpectrosc.212333201327containfalseinformation.AsfortheotherselectionmethodsitcanbeseenthattheEuclideandistancetopredictiononmeancentredspectraldatawasthebestmethodoverallbutthattheEuclideandistanceinPLSspacealsoworkedverywellontheauto-scaledspectraldata.AsacontrastwiththisEuclideandistanceinPLSspacedidnotperformverywellonmeancentredspectraandEuclideandistancetopredictionwasconsiderablymoreinaccuratethanthesamedistanceinPLSspaceonauto-scaledspectra.Table2alsoshowsthatthecorrelationbasedselectionmethodsperformedpoorlyregard-lessofthespectraltreatmentwhichwasalsothecaseforthespectraldistance.AlltheotherselectionmethodsresultedinareductionintheRMSEPthatcouldbedescribedassignicant.HoweverincomparisontothemassiveimprovementobtainedbyusingtheEuclideandistancetopredictiontheimprove-mentthattheseselectionmethodsresultedinseemsinsig-nicant.Afewregressionswerealsodoneforcombinationsofselectionmethods.Byremoving14spectraforeachlayerSizeoflocalsubsetofcalibrationspectraa5122561286432PLS0.008406369966871040672118658211884960.258376089546651041628109154910864470.507725659496011198527131248013594490.7576145910914851158512140934313723731.008555721541980148179914458861462846PCR0.00144857413675341318572138952313935220.25149956213524751285470129043114634250.50158654613684161387384144741815584430.75166451815484261509326148224913193101.0016046871662923150688215078851544900GlobalPCR0.00138657313975451491550167051520156780.25139257513945351507502168445716258130.50141456214324891529470164643516716020.75148454014744921519394157441715455261.0014466871622990148091615349531494888Table1.RMSEPvaluesmgL1obtainedforammoniumonthereducedvalidationdatasetwithLWRusingthespeciedmodelparametersandmeancentredautoscaledspectralpre-processing.AmmoniumAcetateMeancentreAutoscaleMeancentreAutoscaleSectionmethodMinaRelat.bMinaRelat.bMinaRelat.bMinaRelat.bEuclideandistancetoprediction1860.224450.556440.896310.95Euclideandistanceinspectralspace7790.916910.857080.986560.99EuclideandistanceinPLSspace6700.782960.367090.986611.00Absolutedistanceinspectralspace6320.746260.777080.986570.99AbsolutedistanceinPLSspace6390.754110.517261.006611.00MahalanobisdistanceinPLSspace6230.735100.636900.956611.00Correlationinspectralspace7690.907620.946890.956520.99CorrelationinPLSspace8420.988030.997180.996611.00Regressionresidual8561.008091.006180.856570.99aLowestRMSEPmgL1obtainedwiththeselectionmethodinquestionbLowestRMSEPobtainedrelativetotheRMSEPoftheglobalmodelTable2.RMSEPvaluesmgL1obtainedwithML-PLSforammoniumonthereducedvalidationdatasetandforacetateonthefullvalidationdatasetusingthespecieddistancemeasurementstodeterminethelocalsubsetsofmodelregressionspectrawithmeancentringorautoscalingasspectralpre-processing. 28QuantitativeMeasurementsofADProcessParametersbasedontheEuclideandistancetopredictionandfourbasedontheMahalanobisdistanceinPLSspaceonmeancentredspectraaminimumRMSEPof337mgL1wasobtained.Onauto-scaledspectrathecombinationof14spectraremovedaccordingtotheEuclideandistancetopredictionandfourspectraremovedaccordingtotheMahalanobisdistanceinPLSspacethecombinationofeightspectraremovedaccordingtotheEuclideandistancetopredictionandtenaccordingtotheEuclideandistanceinPLSspaceresultedinRMSEPvaluesof363mgL1and366mgL1respectively.ThusfromthesecombinationsonlytheRMSEPof363mgL1waslowerthanthebetteroneofthetwoselectionmethodsusedalone.ThegeneralimpressionoftheseresultswasaswasearlierstatedfortheregressionparametersforLWRthatselectionmethodsshouldbescreenedthoroughlyandsystematicallyinordertoestablishthebestoneorthebestcombinationofmethods.EvaluationofthestepwiseprocedureinML-PLSTheresultsaccountedforuptothispointcanbesummarisedbyaRMSEPof186mgL1withML-PLSandaRMSEPof249mgL1forLWR.Bothofthesearefourtovetimeslowerthanwhatwasachievedwithaglobalmodel.ItcouldalsobearguedthatML-PLSisthemorepowerfulmethod.Howeveritisimportanttorecognisethattheselectionmethodofauto-scaleddistanceinPCspacewasnotevaluatedforML-PLSthatnoattemptsweremadetofullyoptimisethenumberofspectrausedinthelocalsubsetofregressionspectraforLWRandthatthesmalldeviationinaccuracybetweenthesetwomethodscouldbeacoincidence.Hencethistypeofcomparisonisatbestofsecondaryinterest.HoweverithasbeenpointedoutpreviouslythatthekeyfeatureofML-PLSisthestepwisereductionofthelocalsubsetofcalibrationspectra.Thisstatementwasthereforeevaluatedbytworeadilyimplementablemethods.TherstwastocomputetheRMSEPfordifferentsizesofthelocalsubsetofcalibrationspectrawithatwo-layerML-PLSdesign.ThesecondwastostudytheRMSEPobtainedonthelastlayerafter800spectrahadbeenremovedfromthefullregressionsetwithdifferentnumbersoflayers.InbothcasestheEuclideandistancetopredictionincombinationwithmeancentredspectraldatawereusedintheregressionoftheML-PLSmodel.TheresultsfromthersttestcanbeseeninFigure1.ThisgurealsocontainstheRMSEPvaluesobtainedforthethreeregressionmethodsusedinLWRfora0.AnumberofobservationscanbemadeinFigure1.OneisthatthelowestRMSEPobtainedwithatwo-layerML-PLSisroughlythreetimeshigherthanforthe56-layerML-PLSpreviouslyaccountedfor.ThisisastrongindicationthatthestepwisereductionofthelocalcalibrationsubsetisessentialfortheperformanceofML-PLS.Anotheristhattheauto-scaleddistanceinPCspaceisamorepowerfulselectioncriterioninaone-stepselectionofthelocalcalibrationsubsetthantheEuclideandistancetopredictionforthesedatasinceLWRperformedsignicantlybetterthanML-PLSinthistask.ThethirdobservationisthatPCRappearstobeamorereliableregressionmethodformostsizesofthelocalcalibrationsubsetsthanPLS.Thisisinaccord-ancewiththeresultsforLWRpreviouslyaccountedforwherethelowestRMSEPwasobtainedwithaPCRmodelalthoughatadifferenta-value.Theresultsofthe2ndevaluationoftheusefulnessofthestepwiseprocedurei.e.thereductionoftheregressiondataby800spectraindifferentnumberoflayersareshowninTable3.ThesedataclearlyshowsthatthenumberoflayersusedwillaffecttheaccuracyoftheML-PLSmodel.Ifthestepwiseprocedureisessentialtotheperformanceitwillperhapsnotbesurprisingthattoofewlayersresultinlowerpredictionaccuracy.Howevertheresultsalsoindicatethatthereisapenaltyinusingadesignwithtoomanylayers.Althoughthisparticularbehaviourisnotfullyunderstooditisanindicationthatthedesigncanbeoptimisednotonlyintermsofminimisingthenumberoflayersbutalsointermsofmeas-urementaccuracy.OptimisationofthedesigninML-PLSSincetheresultsoftheevaluationoftheusefulnessofthestepwiseprocedureindicatedthatthereisanoptimallayer45050055060065070075080085090002004006008001000RMSEPmgL-1SizeoflocalsubsetofcalibrationspectraFigure1.RMSEPvaluesobtainedfordifferentsizesofthelocalsubsetofcalibrationspectrafora2layermulti-layerpartialleast-squaresML-PLSusingtheEuclideandistancetopredictionasselectionmethodlocallyweightedregressionLWRwithpartialleast-squaresPLSregressionprincipalcomponentregressionPCRandglobalPCRallLWRfora0.layers35917334181161401801RMSEPonlastlayermgL1452519441233267403396396396395Table3.TheRMSEPobtainedforammoniumwithmodelsregressedon408spectrainthelocalsubsetonthelastlayerfordifferentnumberoflayersintheML-PLSdesign. J.DahlbackaandT.LillhongaJ.NearInfraredSpectrosc.212333201329structureacrudeattemptwasmadetomanuallyfindthis.ThiswasdonebystudyingtheresponseofRMSEPtodifferentstepsizesindifferentregionsinthedesign.Thebasicprinciplewastoremoveaspeciednumberofregressionspectrainoneormorestepsandtoevaluatehowmanystepstousetoremovethespeciednumberofspectra.Table4summarisestheresultsofthisevaluation.InthetabletheRMSEPisgivenasrelativetothemaximalstepsizeusedintheremovalinterval.AccordingtoTable4thereislittleornoeffectontheaccuracybyusingmorethanonesteptoremovetherst400spectra.Hence400spectrawasseenastheoptimalstepsizetoremovetherst400spectrafromlayer1tolayer2.Intheintervalof400800removedspectrathestepsizestartstogetrelevant.AscanbeseeninTable4toofewortoomanylayersinthisintervalwilldecreasetheaccuracy.AccordingtoTable4theoptimalstepsizeis25spectraperlayer.Howeveriftheinterval400600removedspectraisexaminedtheoptimalstepsizeinthisintervalis50.Thusastepsizeof50shouldbeusedtoremovethespectrabetween400and600calibrationspectraremoved.Intheintervalof600800removedspectrathechoiceforstepsizecouldeitherbe25or10spectraperlayer.Howeverbetween600and700removedspectra25spectrawasslightlymoreaccurateaswellasamoreeffectivestepsize.Thusintheintervalbetween600and700removedspectra25spectrashouldberemovedfromlayertolayer.AccordingtoTable4theremovalof20spectraforeachlayeristhemostaccuratewaytoremovespectraintheintervalof700840spectraremoved.Thiswasthelaststepsizethatwasoptimised.Fortheremaininglayersastepsizeoftenwaschosenwithoutanyfurtherregardtowhetherornotanotherstepsizewouldhavebeenbetter.WiththisnowestablisheddesigntheRMSEPonlayers26and30was187mgL1and176mgL1respectively.Consequentlythisoptimisationattemptdidnotincreasethemeasurementaccuracycomparedtothe56layerdesignusedpreviouslybutthenumberoflayersthatwasneededtoreachthesameaccu-racywasreducedto26.Figure2showshowtheRMSEPvaluechangesthroughthelayersofthisdesigntogetherwiththenumberofcalibrationspectrathathasbeenremovedbeforeeachlayer.Figure3showsthescatterplotsforatheglobalPLSmodelbthemostaccuratemodelobtainedwithLWRandclayer30intheoptimisedML-PLSmodel.Basedonthesescatterplotsnoneofthemodelsisimpairedbyanobviousslopeoroffseterror.OutlierstrategiesfortheammoniummeasurementSincetheresultsaccountedforonthereducedammoniumvalidationsetindicatedthattheuseoflocalmodelscouldreducetheRMSEPdrasticallyitwasdeemedtobeofgreatinteresttoinvestigatethesourceoftheapparentoutliersinthefullvalidationsetforammonium.Inthiscontextthedeni-tionofanoutlierisalsoimportant.Grubbs28suggestedthatAnoutlyingobservationoroutlierisonethatappearstodeviatemarkedlyfromothermembersofthesampleinwhichitoccurs.HodgeandAustin29exploredthedenitionsofoutliersfurtherandconcludedthatthereisseeminglynouniversallyaccepteddenitionforoutliers.Asthismaybethecasethetermapparentoutlierisusedherefordatathatfullsthede-nitionbyGrubbsasgivenabovewithoutconsideringtheactualsourceofthedeviation.Insteadanattemptismadetodistin-guishbetweenregressionmethodoutliersandconcentrationStepsize400200100805040252010502Removalinterval04001.001.010.980.981.001.011.011.021.021.021.034008001.001.030.951.110.790.670.640.910.891.011.014006001.001.040.950.970.960.990.980.980.986008001.000.850.830.870.821.031.036007001.000.951.020.961.161.477008501.000.961.001.011.01700840forstepsize20Table4.RMSEPvaluesrelativetotheRMSEPofthelargeststepsizeusedonthereducedammoniumvalidationdatasetfordifferentinter-valsofnumberofcalibrationspectraremovedincomparisonwiththefullammoniumregressiondataset.4005006007008009001000150300450600750900051015202530SpectraremovedRMSEPmgL-1LayerFigure2.TheRMSEPvalueforthereducedammoniumvalida-tiondatasetplottedtogetherwiththenumberofcalibra-tionspectraremovedbeforeeachlayeragainstthelayernumber. 30QuantitativeMeasurementsofADProcessParametersoutliers.Basicallyaconcentrationoutlierisconsideredhereasasampleforwhichthereferencevalueconcentrationiserro-neousandaregressionmethodoutlieranapparentoutlierthatmayormaynotbeanoutlierdependingonwhichmultivariatemethodisusedtodeterminethesamplesconcentration.Thusaregressionmethodoutlieristhecombinedimpactofspectralinformationandofhowitisanalysed.Inthecaseswhereasinglespectrumamongthethreereplicateswaspredictedwithaseeminglypooraccuracyitseemedunlikelythatthiswouldbeaconcentrationoutlierandthetasktocorrectthepredictionappearedstraightforward.Iftheabsolutedifferencebetweenapredictionandthemedianpredictionofthethreereplicatesexceedsavaluerelativetothemedianpredictionofthethreereplicatesthisisanindicationthatthespectraisanoutlierandthepredictionisreplacedbytheaverageofthetwootherpredictions.Intheresultspresentedherethismethodwasusedwithathresholdvalueof10onallthemodels.HoweverthishadaverylimitedimpactontheRMSEPvalue.ForinstancefortheoptimisedML-PLSmodeltheRMSEPwasreducedby30mgL1.Figure4showsthescatterplotsforatheglobalPLSmodelbthebestLWRmodelPCRregression64spectrainthelocalsubsetandauto-scaledspectraandclayer30intheoptimalML-PLSmodel.TheRMSEPofthesethreemodelsare897mgL1489mgL1and646mgL1respectively.HoweverFigure3.ThepredictedconcentrationplottedagainstthereferenceconcentrationonthereducedammoniumvalidationdatasetforatheglobalPLSmodelbthebestLWRmodelfoundandclayer30intheoptimisedML-PLSmodel45linedashed.Figure4.ThepredictedconcentrationplottedagainstthereferenceconcentrationonthefullammoniumvalidationdatasetforatheglobalPLSmodelbthebestLWRmodelfoundlayer30intheoptimisedML-PLSdesignusingctheEuclideandistancetopredictionasselectionmethodandmeancentringasspectralpre-processingdtheEuclideandistanceinPLSspaceasselectionmethodandauto-scaleasspectralpre-processingeacombinationofthesetwoselectionmethodsandfathresholdbasedcombinationofthepredictionsshownindande45linedashed. J.DahlbackaandT.LillhongaJ.NearInfraredSpectrosc.212333201331ascanbeseentheRMSEPoftheML-PLSmodelinparticularaswellaspartlyfortheLWRmodelisheavilyaffectedbythepresenceofafewapparentoutliers.Thequestioniswhethertheseareconcentrationoutliersorwhethersomespectralinformationturnsthemintoregres-sionmethodoutliers.SomeoftheapparentoutliersaresharedbetweenML-PLSandLWR.Thiscouldbeseenasanindica-tionthatthesesamplesareconcentrationoutliers.HoweverinordertofurtherexplorethisassumptionanML-PLSmodelwasregressedwiththeoptimiseddesignbutusingtheEuclideandistanceinPLSspaceandauto-scalingforspectralpre-treatment.ThisselectionmethodperformedwellonthereducedvalidationdatasetforammoniumandtheresultsfromthefullvalidationdatasetareshowninFigure4d.ForthismodelaRMSEPof411mgL1wasobtainedandthedifferencecomparedtothepreviouslyusedML-PLSmodelbecomesapparentwhenFigure4cand4darecompared.AlthoughthepredictionsingeneralseemtobelessaccuratecomparedtotheotherML-PLSmodelusedthismodeldoesnotproducepredictionswithamassiveresidualinthesamewayastheothermodel.Thusitseemsunlikelythatthevali-dationdatasetincludedconcentrationoutlierstoanygreaterextent.ItwasevaluatediftherobustnessoftheEuclideandistanceinPLSspacecouldbecombinedwiththeaccu-racyoftheEuclideandistancetopredictionastheselectionmethod.Thiswasdonebyusingthelatterastheselectionmethodfortherst17layersandthereafterswitchingtotheformerontheremaininglayers.ThisresultedinamodelwithaRMSEPof665mgL1andthepredictionsonthevalidationdatasetareshowninFigure4e.Sincethisattemptdidnotprovidetherobustandaccuratemodelthatwasitsgoalitwasdecidedtotestwhethertherobustnessandaccuracycouldbeobtainedbycomparingthepredictionsofthetwolastmodelsdescribedhere.Aselectionofpredictionswasperformedbystatingthatanypredictionsmadebythelastmodeldescribedcouldnotdivergemorethan1300mgL1fromthepredic-tionsofthemodelthatusedtheEuclideandistanceinPLSspaceasselectionmethod.ThiswayapredictionaccuracydescribedbyaRMSEPof274mgL1wasobtainedFigure4f.Althoughthiscanbeseenasacumbersomemethodtodealwiththeapparentoutliersitclariedthequestionwhetherornottheseoutlierswereconcentrationoutliersorregressionmethodoutliers.LocalcalibrationmodelsforacetateWhereasthelocalcalibrationtechniquesusedshowedgreatpotentialforthemeasurementoftheammoniumconcen-trationthesituationwasquitetheoppositefortheacetateconcentration.TheresultsoftheLWRmodelsareshownasRMSEPvaluesinTable5.SincetheRMSEPoftheglobalPLSmodelwas723mgL1withmeancentringand664mgL1withauto-scalingasthepre-treatmentitcanbeconcludedthatnotasingleLWRmodelwiththe150regressionparametersettingsusedproducedaRMSEPlowerthananyoftheglobalmodels.ThesituationwasnotmuchmoreencouragingwithML-PLSeither.InTable2theresultsoftheevaluationoftheselectionmethodscanbeseen.ItisperhapssymptomaticthatthelargestreductionoftheRMSEPisobtainedwhentheregressionspectraareremovedbasedonthepredic-tionresidual.FurthermorethereductionsinRMSEPwerebasicallyonlyobtainedwhenmeancentringwasusedaspre-processingofthespectra.OntheotherhandtheglobalPLSonauto-scaledspectrawassomewhatmoreaccuratethanthecorrespondingoneonmeancentredspectra.SinceSizeoflocalsubsetofcalibrationspectraa5122561286432PLS0.0077694377490079190483490194610840.2576696675591676092479186792910170.5074899072792674089180286999310020.75752101076489794990511021050127112021.0088695914321349149314261498141414761400PCR0.0010799131025891985897971954103210350.2510919201043889993901967941106110320.501113915105488710329241085995118710700.751389979137110331328109213471203134912191.001492134615601415148413961475139714701383GlobalPCR0.0013451363129612941269120712681193138814110.2513441347128112601234117512561164130014450.5013441329127112411233115613151205140013600.7514041318141812891446124415211337153114121.001568150415621415148413751479141614881401Table5.RMSEPvaluesmgL1obtainedforacetateonthefullvalidationdatasetwithLWRusingthespeciedmodelparametersandmeancentredautoscaledspectralpre-processing. 32QuantitativeMeasurementsofADProcessParameterssomeimprovementsinthemeasurementaccuracywasstillobtainedanumberofdesignsandcombinationsofselec-tionmethodsweretested.ThesetestsindicatedthatahighnumberoflayersseemtoenhancetheaccuracythatusingtheEuclideandistancetothemedianpredictiononpreviouslayershadsomebenetsandthatdecreasingthenumberofPLScomponentsinthemodelastheregressiondatasetdecreasescouldbefavourable.Witha249layerML-PLSinwhichthreespectrawereremovedeverylayerbasedontheEuclideanpredictiontothemedianpredictionofthepreviouslayersandonespectrumbasedonthepredictionresidualandbyusingonlythreecomponentinthemodelsafterlayer200aREMSEPof576mgL1wasobtained.Whetherornotthiswasasignicantreduction21incomparisonwiththeglobalPLSmodelisamatterofdenition.HoweverwhatiscompletelyclearisthatitwasbynomeansassignicantasthereductionsobtainedforammoniumwithLWRandML-PLS.ConclusionsThisstudyaimedatevaluatingthelocalcalibrationtechniquesLWRandML-PLSforquantitativemeasurementsofammoniumandacetateinsamplesfromanADprocessaswellasatfurtherexploringthepreliminaryconceptoftheML-PLSalgorithm.ItwasshownthatforammoniumtheRMSEPofaglobalPLSmodelat897mgL1couldbereducedto489mgL1withLWRandto274mgL1withML-PLS.IncontrasttotheseexcellentresultsnoreductionintheRMSEPvaluewasobtainedforacetatewithLWRandonlyaninsignicantreductionwithML-PLS.Thustwodiametri-callydifferentresultswereobtainedforthetwocomponentsalthoughtheywereinthesameconcentrationrangeandinthesamebackgroundmatrixaswellashadaverysimilarRMSEPfortheglobalPLSmodels.TheconceptofML-PLSwasexploredbymakingpredictionsonareducedvalidationdatasetforammonium.Therstissuethatwasaddressedwasacomparisonofdifferentselectionmethodsfordeter-miningthelocalsubsetofcalibrationdata.InthiscasetheEuclideandistancetopredictiongavethelowestRMSEPat186mgL1onmeancentredspectraldata.HowevertheEuclideandistanceinPLSspacealsoperformedwellresultinginaRMSEPat296mgL1onauto-scaledspectraldata.Theconclusionwhichwasfurtherverifiedonthefullvalidationdatasetforammoniumwasthatthebestselectionmethodsaredifferentfordifferentpre-treatmentmethodsaswellasfordifferentdatasets.TheusefulnessofthestepwiseoriterativeapproachinML-PLSwasevalu-atedrstbyselectingthelocalsubsetofcalibrationspectrainonesinglestepandsecondbycollectingalocalsubsetofthesamesizefordifferentnumbersoflayers.Theresultsshowedthattheaccuracywassignicantlyreducedwhentheselectionwasmadeinonestepbutalsothatusingtoomanystepsorlayersappearstohaveanegativeeffectontheaccu-racy.Thusthendingsindicateapotentialinoptimisingthelayerstructureaswellastheselectionmethodsfordifferentlayers.AcknowledgementsThisworkwasmadepossiblethroughtheMarePurumprojectfundedbytheEuropeanterritorialcooperationprogrammeBotnia-Atlantica.TheauthorsalsowanttothankProfessorPaulGeladiforscienticadvice.References1.A.M.C.DaviesH.V.BritcherJ.G.FranklinS.M.RingandW.F.McClureTheapplicationofFouriertransformedNIRspectratoquantitativeanalysisbycomparisontosimilarityindicesCARNACMikrochim.Acta1611988.2.G.SinnaeveP.DardenneandR.AgneessensGlobalorlocalAchoiceforNIRcalibrationsinanalysesoffor-agequalityJ.NearInfraredSpectrosc.21631994.doi10.1255jnirs.433.T.FearnandA.M.C.DaviesLocally-biasedregressionJ.NearInfraredSpectrosc.114672003.doi10.1255jnirs.3974.J.S.ShenkP.BerzaghiandM.O.WesterhausInvestigationofaLOCALcalibrationprocedurefornearinfraredinstrumentsJ.NearInfraredSpectrosc.52231997.doi10.1255jnirs.1155.E.Zamora-RojasA.Garrido-VaroF.VandenBergJ.E.Guerrero-GinelandD.C.Prez-MarnEvaluationofanewlocalmodellingapproachforlargeandheterogeneousNIRSdatasetsChemometr.Intell.Lab.Syst.1012872010.doi10.1016j.chemo-lab.2010.01.0046.D.Prez-MarnA.Garrido-VaroandJ.E.GuerreroNon-linearregressionmethodsinNIRSquantitativeanalysisTalanta721282007.doi10.1016j.tal-anta.2006.10.0367.X.ShaoX.BianJ.LiuM.ZhangandW.CaiMultivariatecalibrationmethodsinnearinfraredspec-troscopicanalysisAnalMethods216622010.8.W.S.ClevelandRobustlocallyweightedregressionandsmoothingscatterplotsJ.Am.Stat.Assoc.743688291979.doi10.108001621459.1979.104810389.W.S.ClevelandandS.J.DevlinLocallyweightedregressionanapproachtoregressionanalysisbylocalttingJ.Am.Stat.Assoc.834035961988.doi10.108001621459.1988.1047863910.T.NsT.IsakssonandB.KowalskiLocallyweightedregressionandscattercorrectionfornear-infraredreectancedataAnal.Chem.626641990.doi10.1021ac00206a00311.T.NsandT.IsakssonLocallyweightedregressionindiffusenear-infraredtransmittance J.DahlbackaandT.LillhongaJ.NearInfraredSpectrosc.212333201333spectroscopyAppl.Spectrsoc.461341992.doi10.1366000370292444434412.Z.WangT.IsakssonandB.R.KowalskiNewapproachfordistancemeasurementinlocallyweightedregressionAnal.Chem.6622491994.doi10.1021ac00074a01213.A.H.AastveitandP.MarumNear-infraredreectancespectroscopydifferentstrategiesforlocalcalibrationsinanalysesofforagequalityAppl.Spectrosc.474631993.doi10.1366000370293433491214.J.DahlbackaandT.LillhongaMoisturemeasurementintimberutilisingamulti-layerpartialleastsquarescalibrationapproachJ.NearInfraredSpectrosc.184252010.doi10.1255jnirs.90615.M.MadsenJ.B.Holm-NielsenandK.H.EsbensenMonitoringofanaerobicdigestionprocessesAreviewperspectiveRenew.Sustain.EnergyRev.1531412011.doi10.1016j.rser.2011.04.02616.M.HanssonA.NordbergandB.MathisenOn-lineNIRmonitoringduringanaerobictreatmentofmunicipalsolidwasteWaterSci.Technol.48492003.17.M.HanssonA.NordbergI.SundhandB.MathisenEarlywarningofdisturbancesinalaboratory-scaleMSWbiogasprocessWaterSci.Technol.5102552002.18.C.J.LomborgJ.B.Holm-NielsenP.Oleskowicz-PopielandK.H.EsbensenNearinfraredandacousticchemometricsmonitoringofvolatilefattyacidsanddrymatterduringco-digestionofmanureandmaizesilageBioresour.Technol.100517112009.doi10.1016j.biortech.2008.09.04319.J.B.Holm-NielsenH.AndreeH.LindorferandK.H.EsbensenTransexiveembeddednearinfraredmonitoringforkeyprocessintermediatesinanaerobicdigestionbiogasproductionJ.NearInfraredSpectrosc.151232007.doi10.1255jnirs.71920.D.WolfH.vonCansteinandC.SchrderOptimisationofbiogasproductionbyinfraredspectroscopy-basedprocesscontrolJ.Nat.GasSci.Eng.36252011.doi10.1016j.jngse.2011.07.00621.A.J.WardP.J.HobbsP.J.HollimanandD.L.JonesOptimisationoftheanaerobicdigestionofagriculturalresourcesBioresour.Technol.991779282008.doi10.1016j.biortech.2008.02.04422.J.DahlbackaT.LillhongaandM.DringDesignedorthogonalsamplespikingbasedcalibrationsforquan-titativeliquidphasemeasurementswithnearinfraredspectroscopyinananaerobicdigestionprocessJ.NearInfraredSpectrosc.21112013.doi10.1255jnirs.103223.WaterQuality.DeterminationofAmmoniumNitrogen.MethodbyFlowAnalysisCFAandFIAandSpectrometricDetectionSFS-ENISO11732.FinnishStandardsAssociationSFSHelsinkiFinland2005.24.M.-H.YangandY.-M.ChoongArapidchromatographicmethodfordirectdeterminationofshort-chainC2C12volatileorganicacidsinfoodsFoodChem.751012001.doi10.1016S0308-81460100211-425.T.NsT.IsakssonT.FearnandT.DaviesAUserFriendlyGuidetoMultivariateCalibrationandClassication.NIRPublicationsChichesterUK2002.26.A.SavitzkyandM.J.E.GolaySmoothinganddifferentiationofdatabysimpliedleastsquaresproceduresAnal.Chem.36816271964.doi10.1021ac60214a04727.B.M.WiseN.B.GallagherR.BroJ.M.ShaverW.WindigandR.S.KochChemometricstutorialforPLS_ToolboxandSolo.EigenvectorResearchInc.3905WestEaglerockDriveWenatcheeWA98801USA200628.F.GrubbsProceduresfordetectingoutlyingobserva-tionsinsamplesTechnometrics11111969.doi10.108000401706.1969.1049065729.V.J.HodgeandJ.AustinAsurveyofoutlierdetectionmethodologiesArtif.Intell.Rev.22852004.doi10.1007s10462-004-4304-y JOURNALOFNEARINFRAREDSPECTROSCOPY35ISSN0967-0335IMPublicationsLLP2013doi10.1255jnirs.1034AllrightsreservedTheidentificationandqualitycontrolofdairyproductsiscarriedoutworldwidebymeasurementofthefatcaseinandwheyproteinssugarstotalsolidssomaticcellstypicalbacteriaandoftheotherregulatedindicatorsusingmulti-steplaboratorymethodswiththeapplicationofachroma-tographychemicalmetrologynuclearmagneticresonanceorFourier-transformspectroscopycombinedwithchemo-metriccalibration.1NearinfraredNIRradiationscatteredorabsorbediswidelyappliedasaquantitativespectroscopytechniquetoassessmilkcomposition2asitisanefcientrelativelysimpleandlowcostmethod.Aportabletoolforoper-ativedeterminationofqualityindicatorsforliquidmilkwouldclearlyhavevalue.AttemptstouseashortwaveNIRSWNIRspectrometerwithalinearsiliconCCDarrayforsimultaneousdeterminationoftheconcentrationsofseveralcomponentsinmilkarewellknown.35HoweveravailablereportsoftheShort-wavenearinfraredspectrometryofbackscatteringandtransmissionoflightbymilkformulti-componentanalysisAndreyV.KalininViktorN.KrasheninnikovandVladimirM.KrivtsunInstituteforSpectroscopyoftheRussianAcademyofSciencesTroitskRussia.E-mailkalininisan.troitsk.ruQualityandauthenticityofmilkareusuallydenedbythecontentandcompositionoffatproteinstotalsolidssomaticcellsandotherestablishedindicators.AlongwithtraditionalreferencemethodsforquantitativeanalysistheuseofnearinfraredNIRspectrometersandmultivariatemodellingisbecomingmorewidelyutilised.ItwouldbeusefultoapplyaspectrometerwithalinearsiliconCCDarrayasarapidportableandrathercheapsupervisingdevice.Howevermilkisamultiphasedispersesystemwhichincludesaroughdis-persephaseofsomaticcellsandfattyglobulesathinphaseofcaseinandwheyproteinsparticlesandthemolecularsolutionsoflactosesaltsvitaminsetc.inwater.ThescatteringofincidentlightbyparticlesofmilkaffectthesignalofanNIRspectrometersignicantly.Thecalibrationofaspectrometercanbeimprovedbyconsideringthedistinctionbetweenscatteringbyfattyglobulesandcaseinmicelles.Forthispurposeitisattractivetousetheanisotropyofscatteringbyroughparticlesinparticularbyfattyglobules.Thecompara-tiveresultsforcalibrationmodelsconstructedwithtransmissionandbackscatteringspectraofmilkarepresentedinthispaper.Thespectrawereacquiredwithanewtwo-channelshortwaveNIRspectrometer.Restoreddrinkingmilkhasbeenmodelledwithasetofcalibrationsamplessincethedifferenceintheopticalpropertiesduetotheglobulesandthemicellesissignicantandalsobecausethereliablequalitycontrolofdrinkingmilkusingasimplespectrometerisarelevantpracticalproblem.PLSmodelswereconstructedforthepredictionoffatandproteinandalsoofthefat-freetotalsolidswhichserveasanindicatorofauthenticityfordrinkingmilk.Theresultsobtaineddemonstratethatalargereductionsinerrorsispossible.TheSEPvalueofnotmorethan0.08wtwasachievedwhenpredictingthefatcontentofmilkusingacalibrationbasedonthedifferencespectraoftransmissionandbackscatteringcomparedwithaSEPof0.21wtforthemodelusingtransmissionspectraonly.ForproteindeterminationtheSEPwasreducedfrom0.25wtto0.12wtifthecalibrationwasbasedonthespectraofbackscatteringratherthanontransmission.KeywordsnearinfraredspectrometrylightscatteringanisotropymilkfattyglobulecaseinmicelleprojectionsonlatentstructuresIntroductionA.V.KalininV.N.KrasheninnikovandV.M.KrivtsunJ.NearInfraredSpectrosc.2135412013Received2June2012Revised30October2012Accepted29November2012Publication21January2013 36Multi-ComponentAnalysisofMilkabove-mentionedmilkanalysershavedemonstratedthattheylackreliabilityandprecisionincomponentdeterminationseeReference6forexampleincomparisonwithofcialmethodsseeReference7forexample.WeagreewiththestatementsbyCzarenaetal.8andWalstra9thatthisisduetothedomi-nanceofscatteringoverabsorptioninliquidmilkandthusthesubstantialinuenceofparticlesizeonthespectralvariationandchemometriccalibration.Milkisamultiphasedispersesystem.10Theroughdispersephaseconsistsofsomaticcellsbacteriaandfattyglobulesthethinphaseisthepopulationsofcaseinmicellesandaggregatesofwheyproteinandtheremainderisadispersivemediuminvolvingmolecularsolu-tionsofdisaccharidelactosesaltsvitaminsetc.inwater.OpticalpropertiesofmilkintheSWNIRspectralrangearecharacterisedbythepredominantscatteringoflightbytwotypesofparticlesfattyglobulesofaveragesize0.510mandcaseinmicelles0.10.2m.Considerabledistinctioninthesizeandrefractiveindicesofmilkparticlesleadstovarioustypesofscatteringdiffusivebymicellesandstronglyaniso-tropicbyfattyglobules.ThereforethecalibrationofaNIRspectrometershouldtakeintoaccountpossibledifferencesinparticles.Earlierwemanagedtoconrmthepossibilityofseparatelydeterminingcaseinmicellesandaggregatesoflacto-globulinbytheapplicationoftransmissionandtrans-ectionmodesforacquiringNIRspectra.11Howeveranyinfor-mationaboutmodellingthesizevariationinfattyglobulesindependentlyfromfattinessinrawmilkisabsentasfarasweknow.Thedifferenceinsizesandrefractiveindicesoffattyglobulesandcaseinmicellesarelargeenoughtodetecttheircorrelationwiththeextinctionspectraofmilksoitisconven-ienttousereconstitutedmilkforthecalibrationsamplesetandthesimpleprocedureoftheirpreparationbymixingtheparentmaterials.Thepurposeofthisstudywastoinvestigatethecompara-tivecharacteristicsofprojectionstolatentstructuresPLSmodelsbasedonSWNIRspectraofbackscatteringandtrans-missionofreconstitutedmilkforpredictingtheconcentrationoffatandtotalprotein.Anadditionalobjectivewastoexplorethepossibilityofpredictingthevalueoffat-freetotalsolidsFFTSwhichistheofficialindicatorofauthenticityfordrinkingmilk.ThespectrawereobtainedwithanewportabledualSWNIRspectrometer.MaterialsandmethodsSamplesandreferencedataThecalibrationandtestsamplesetsnumbering36and40respectivelywerepreparedbymixingvaryingweightsofinitialmaterials10creamsuppliedbyKampinaLtdMoscowRegionRussiaskimmedmilkreceivedfromtheexperi-mentalfarmoftheRussianScienticResearchInstituteofDairyIndustryVNIMIMoscowmilklactoseobtainedfromRi_gasPiensaimnieksLtdLatviaanddistilledwater.SetsofcalibrationandtestsampleswerepreparedinthefollowingwaythefatcontenttotalproteinlactoseandFFTSintheinitialmaterialsweremeasuredrstusingtherefer-encemethodsasdescribedbyKalininetal.12intheMilklabo-ratoryVNIMIandthenthemassfractionsoffattotalproteinandFFTSforthesamplesinthesetswererandomlyassignedtoavoidproportionalvariationofcomponentsconcentra-tioninsetsandtheweightofinitialmaterialsfromthegivenvalueswerecalculated.TherangesofmassfractionvariationforfatproteinlactoseandFFTSselectedweretypicalfordrinkingmilkandwere2.04.52.23.254.85.0and7.610.4inwtunitsrespectively.Thelactosecontentinthecalibra-tionsampleswassetataleveltogivesamplesthevaluesoftherefractiveindexandlightscatteringpropertiestypicalformilk.WeighingwasexecutedscalesMA35manufacturedbySartoriusGttingenGermanywithanerrorof3mginglassbottleswithacapacityof12mL.Theweightofrawmaterialsreducedtoroomtemperatureweresubsequentlyplacedintostacksandtoppedupwithwatertoaweightof10g.Thesesampleswerewarmeduptoatemperatureof40Chomo-genisedwithanultrasonicthermostatSapphire0.5-2SapphireLtdMoscowRussiafor10minthenleftforabout5htoallowthestabilisationofcolloidalstructuresbeforeanymeasurementswereacquired.ThespectrometryofmilksamplesThedifferencesbetweentheanisotropyandtheintensityoflightscatteredbythemicellesandtheglobulescanbeusedtoimprovetheregressionmodelbyrecordingtheintensityofbackscatteringandtransmissionspectrasimultaneously.Todothiswehaveproducedatwo-channelspectrometerdetailsbelowforsimultaneouslyacquiringtheextinctionspectraat0degreesand150degreestotheaxisoftheilluminatingbeam.TwogratingspectrometersofthenameddevicewereequippedwithtwolinearCCDswithaspectralsensitivitytolightandawavelengthrangefrom800nmto1065nm.Thespectrareectboththespectralangularcharacteristicsoflightscatteringbyparticlesofvarioussizesandspectralfeaturesofmolec-ularabsorptionofcomponentsforexamplethewell-knownovertoneofbasicfatabsorptionnearwavelength930nmorthewaterabsorptionbandnear970nmoracombinationabsorptionbandofproteinandwaternear1020nm13.Theknownvaluesoftheabsorptionandscatteringcoefcientsofmilkaregroundsforchoosingtheshort-waveborderofaspectralrangenear800nm.AccordingtoYaroslavskyetal.14theabsorptioncoefcientofmilkis2030timeslessthanthescatteringcoefcientforthelightatthiswavelength.Despitehavinglowabsorptionitslevelissufcientlyhightodetecttheovertonesandcombinationalabsorptionoffatsproteinsandwaterasisevidentfromthespectraofloadingsandregres-sioncoefcientsreportedbySasicandOzaki.3OtherfeaturesoflightscatteringbymilkfollowfromintensityangulardependencieswhichwerecalculatedaccordingtotheMietheory15atwavelength800nmfordielec-tricsphereswiththerealrefractionindex1.46asforfatforvariousdiametersfrom1to2.75whicharecharacter-isticforfattyglobulesofhomogenisedcowmilkinawaterenvironmentseeFigure1.Thespeciedincreaseinsphere A.V.KalininV.N.KrasheninnikovandV.M.KrivtsunJ.NearInfraredSpectrosc.213541201337diameterwasaccompaniedbyanincreaseintheintensityofforwardscatteringatanangle0degreestoanincidentlightdirectionofapproximately27000timeswhereasthelevelofbackscatteringatanangle150degrees170degreeswasnotrelatedtoanincreaseindiameter.ThisnumericalresultfromtheMiesolutionisforsinglescatteringbutcanpointtoaqualitativeeffectforthemulti-scatteringinwholemilk.AnisotropyofradiationscatteredbytheroughmilkphasefattyglobulesinthewholemilksampleisdemonstratedclearlyinourphotographspresentedasFigure2.Theabove-mentionedpropertiesofmilkallowustoassumethatSWNIRspectrometrycanbeusefulfordiscriminatingbetweenthedifferenttypesofscatteringduringthethinandroughphasesofmilk.Thecontentofthethinphasecaseinmicellesandparticlesofwheyproteinsshouldberevealedwiththeuseofbackscatteringwhichisaccompaniedbyminimumscatteringoflargeparticles.Incontrastthecontentoftheroughfractionfattyglobulesandsomaticcellsprob-ablyshouldbedeterminedusingthetransmitteduxwhichincludesmaximumscatteringbytheroughparticles.Theopticalschemeofthetwo-channelspectrometercreatedatourInstituteforSpectroscopyTroitskCityMoscowispresentedasFigure3.Distinctivefeaturesofthemethodandthetwo-channelspectrometerwerepatentedinRussia.16.Thecuvetteusedtoholdthesampleswasaglasschemicaltubewitha10mminnerdiameter.Theinletsizeofthetrans-missionspectrometerwas0.8mmthediameterofback-scatteringopticfibrewas0.2mm.Thedistancefromthesampletotheinletoftheopticbrewaschosenas2.3mmsothatthebackscatteringsignalfromthewholemilkandfromthereferencesampleanAl2O5ceramicplatewereofthesamescale.Boththespectrometershadalineardispersionof0.11mmnm1andthespectralresolutionof7nm.TheCCDscontained3648pixels8200insize.Whenacquiringspectraabackgroundcontrolandareferencespectrumwererecordedevery6min.Thelogarithmoftheratioofthereferencespec-trumtoeachmilksamplespectrumwasinsertedintotheself-madeprogramISCAP12togetherwiththedataofcomponentcontentandthesamplename.Despitethefactthatresultinglogarithmsforscatteringandtransmissionhadthedimensionofabsorbanceunitstheywerephysicallydifferentfromtheopticaldensityduetothepresenceofthereferencespectrumthereforetheseweredesignatedasarbitraryunitsa.u..Thetimerequiredtopredictvecomponentsinmilkwasnomorethan1min.Thespectrometerwassmallwithafootprintofonly20cm27cmandweighed5kgwithoutthelaptop.ThenegativelogarithmsoftransmissionandbackscatteringspectraexpressedinabsorbancearbitraryunitswereusedFigure1.Simulatedangulardistributionofsinglescatter-ingintensityforlightatthewavelength800nmbydielectricsphereatrefractiveindex1.46thespherediameterwasvariedfrom1thelowercurvenearangle0degreesto2.75theuppercurvenearangle0degrees.abFigure2.Photostopviewofglass4cmdiameterofamilksurfacewithscatteringauraofalaserbeamwavelength632nmwhichilluminatedtheglassfromtheleftadiffusescatteringinmilkthatwasskimmedandenrichedwithcaseinmicellesto5.5wtoftotalproteinbanisotropicscatteringinwholemilkcontaining3.5wtoffatand2.1wtoftotalprotein. 38Multi-ComponentAnalysisofMilktobuildandtotestthePLSmodels.Spectraof10oftheseexamplesarepresentedinFigure.4.ConstructionandtestingofcalibrationmodelsCalibrationmodelshavebeenexecutedusingthePLSregressionmethod17usingtheself-madeprogramISCAP.12Thespectraandmassfractionsofcomponentsforacalibra-tionsamplesetwereenteredintotheISCAPprogramandthencalibrationmodelswereconstructedforeachdenedcompo-nentusingrawdataandsomepre-treatmentsonthespectraPreliminaryprocessingofspectrabydifferentiationremovalofthegeneralpartofthebaselineorcorrectionofmulti-scatteringdidnotusuallyleadtoimprovementsinthemodelquality.PreferablythereforethesmoothedSavitzkyGolayalgorithmpolynomialdegree3thenumberofpoints20andrawspectrawereused.CombinationsoftransmissionandbackscatteringspectrawereproducedintwowaysrstthemodelsnamedCombwereobtainedbyincludingthebackscatteringspectrumforeachsampleonthewavelengthaxisafterthetransmissionspectrumadding300nmtoeachwavelengthforthesamesampleasitisshownatFigure4andsecondmodelsnamedDiffwereconstructedusingthespectraofdifferencesbetweenthenegativelogarithmsfromthespectraoftransmissionandofbackscatteringforeachsample.ForTransmissionthemodelswerebasedonnegativelogarithmsoftransmissionandforScatteringnamedmodelsusingthenegativelogarithmsofscattering.Rawmeantuntreatedlog1RspectraRbthespectrawithremovedbaselineCombDer-the1stdifferentialofthecombinedspectraSm-smoothedspectrausingtheSavitzkyGolayalgorithm.ThesamplesforwhichtheMahalanobisdistanceofmorethanfouraccordingto18wereexcludedfromthecalibrationandinthenamesofthecorrespondingmodelsthenumberofexcludedsamples1and3havebeennoted.Cross-validationparameterswerecalculatedusingtheleave-one-outtechnique.12Thequalityofmodelswasestimatedfromthecrossvalida-tionparametersTable1wherenisthenumberoflatentvari-ablesinthePLSmodelwhichaccountedforthedeterminationofagivencomponentr2cvisthesquaredmultiplecorrelationcoefcientamountofthecomponentcontentexplainedofcrossvalidationandRMSECVistheroot-meansquareerrorofcrossvalidation.Thebestmodelsforeachcomponentwereselectedfortestingagainstthetestsampleset.Thevaluesfortheroot-meansquareerrorofpredictionRMSEPandforthesquaredmultiplecorrelationcoefcientofthepredictionfortestsamplesr2pwereextracted.DenitionsofcalibrationparameterswereasdescribedbyWoldetal.17Thesevaluesservetoillustratethetypeofpredictedperformancethatmaybepossibleusingthistechniqueonliquidmilksamples.ResultsanddiscussionMorethan20modelsfordeterminationofthefatproteinandFFTScontentwhichdifferedduetothepreprocessingFigure3.SchemeofSWNIRtwo-channelspectrometer1lightsource2objectivewithcut-oflter3test-tubewiththemilksam-ple4opticalbreofbackscatteringchannel5transmissionandbackscatteringspectrometers6computer.Figure4.Thecombinationoftransmissionontheleftandbackscatteringontherightspectraatwavelengthregion8001060nmfor10milksamplesP1-P29arethenumbersofthesamples. A.V.KalininV.N.KrasheninnikovandV.M.KrivtsunJ.NearInfraredSpectrosc.213541201339ofthespectraandsomeexcludedsampleswerechosenandtested.Nineofthesemodelswiththecrossvalidationparametersnr2cvandRMSECVandtestingparametersr2pandRMSEPfortheselectedmodelsforfatandproteinpredictionarepresentedinTable1.ThemassfractionofFFTSisaderivativevaluebecauseitisdenedbasicallybythesumoftheproteinandlactoseandwithotherindicatorsitisusedasanindicatorofauthenticity.RussianFederalLawrequiresthattheallowableerrorforthepredictionofmoisturebytheexpressedmethodforTechnicalRegulationsonMilkandDairyProducts19is0.4wt.InallmodelsdevelopedinthisstudyforFFTStheRMSEPvaluedidnotexceeded0.3wtirrespectiveofthetypeofspectraandthepreprocessingmethodused.ThatresultmeansthatitwillbepossibletoapplythisNIRmethodforFFTSpredictionforpracticaltests.ThetypicalspectrumofregressioncoefcientsforthebestfatpredictionmodelseeFigure5ahasconsiderablevaluesatthosewavelengthswhichwereassignedbySasicandOzaki3toovertonesoffatabsorptionbands930nm1020nmand1040nm.Ontheotherhandtheformofregressioncoef-cientspectrumfortheproteinmodelseeFigure5bwasverydifferentinappearancefromthecurveforfatseeFigure5a.ItshouldbenotedthatthespectrumofloadingsandregressioncoefcientsofthekindpresentedinFigure5bwasobtainedandinterpretedbyBogomolovetal.20topredictthemilkprotein.Figures5aandbshowthatcorrelationmodelsforfatandproteinarebasedoninformationfoundatdifferentregionsofthespectrum.Conclusions1.UseofdifferencespectraoftransmissionandbackscatteringintheSWNIRregioncanincreasetheprecisionoffatpredictionsinsamplesofdrinkingmilk.nr2cvr2pRMSECVwtRMSEPwtModelnameFatProteinFatProteinFatProteinFatProteinFatProteinCombRaw550.840.850.830.790.160.180.190.19CombDer520.840.850.790.720.160.220.130.21CombRb330.850.870.810.860.120.270.130.25DiffRaw550.960.840.920.800.080.160.080.17DiffSm430.990.880.980.840.070.110.090.14TransmissionRb440.870.830.800.730.170.180.210.25TransmissionDer680.860.840.790.700.190.240.270.21ScatteringRb1140.820.870.790.870.170.110.190.12ScatteringRb3230.870.820.720.820.170.100.180.12n-thenumberofPLSlatentvariablesr2cv-squaredmultiplecorrelationcoefcientforcross-validationr2p-squaredmultiplecorrelationcoefcientforthepredictionoftestsamplesetRMSECV-therootmeansquareerrorofcross-validationRMSEP-therootmeansquareerrorofpredictionCombscattercombinedwithtransmissionspectraRawuntreatedspectramodelCombDermadeas1stderivativefromcombinedandsmoothedspec-tramodelCombRbmadeusingspectrawithbaselineremovedDiffdifferencebetweentransmissionandscatterspectramodelsScatteringRb1andScatteringRb3whichdifferedduetothenumberofexcludedcalibrationsamplesoutliers1or3respectively.Thebestvaluesoftheparametersareinboldtype.abFigure5.ThespectraofregressioncoefcientsforthecalibrationmodelDiffRawaforfatatvePLSlatentvariablesbforproteintwoPLSlatentvariablesX-axisdigitisedinnanometres.Table.1.ParametersofninePLScalibrationmodels. 40Multi-ComponentAnalysisofMilk2.Subtractionofbackscatteringspectrafromtransmissionspectrahasasignicantimpactonreducingthepredictionerrorduetoparticlesize.3.Theprecisionandreliabilityofproteinpredictioncanbeincreasedbyusingbackscatteringspectracomparedwithapredictionbasedonthetransmissionspectraalone.Toobtaintheprecisionnecessaryforacommercialappli-cationitwillbenecessarytouseaconsiderablylargernumberofcalibrationsamples.Thiswasconfirmedbythefactthatparametersofmodelsimprovedconsiderablywhenbothcalibrationandtestingsetsampleswereusedforcalibration.Theimprovementinthecalibrationmodelconstructedondifferencesoftransmissionandbackscatteringspectratesti-espossiblytoabiggerroleofscatteringbycaseinmicellesthanpredictedfromthetheoryofsinglescatteringandisamanifestationofmulti-scatteringintheconcentrateddisper-sions.AcknowledgementTheauthorsareverygladtoexpressthankstoMrsHelenYurovatheChiefoftheMilLaboratoryVNIMIMoscow.Thespectrometeranditsoperationforthepredictionofmilkqualityindicatorswereshownatthe63rdInternationalIdeasInventionsNewProductsExhibition2730October2011inNurembergGermanywhereitwasawardedasilvermedalandaspecialprizefromtheKoreanAssociationofInventors.References1.R.KarouiandJ.DeBaerdemaekerAreviewoftheanalyticalmethodscoupledwithchemometrictoolsforthedeterminationofthequalityandidentityofdairyproductsFoodChem.1022007621.doi10.1016j.foodchem.2006.05.0422.B.AernoutsE.PolshinJ.LammertynandW.SaeysVisibleandnear-infraredspectroscopicanalysisofrawmilkforcowhealthmonitoringreectanceortransmit-tanceJ.DairySci.941153152011.doi10.3168jds.2011-43543.S.SasicandY.OzakiShort-wavenear-infraredspec-troscopyofbiologicaluids.Quantitativeanalysisoffatproteinandlactoseinrawmilkbypartialleast-squaresregressionandbandassignmentAnal.Chem.73112001.doi10.1021ac00104314.Y.A.WooY.TerazawaJ.Y.ChenC.IyoF.TeradaandS.KawanoDevelopmentofanewmeasurementunitMilkSpec-1forrapiddeterminationoffatlactoseandproteininrawmilkusingnear-infraredtransmittancespectroscopyAppl.Spectrosc.5652002.5.M.KawasakiS.KawamuraM.TsukaharaS.MoritaM.KomiyaandM.NatsugaNear-infraredspectroscopicsensingsystemforon-linemilkqualityassessmentinamilkingrobotComput.Electron.Agric.632008.6.K.SpitzerR.KnnemeyerM.WoolfordandR.Claycomb.On-linemilkspectrometryAnalysisofbovinemilkcom-positionintheProceedingsofSPIETheInternationalSocietyforOpticalEngineeringPartIIArt.No.112pp.6987072005.7.AssociationofOfcialAnalyticalChemistsinOfcialMethodsofAnalysis.15thEdn.AOAC.ArlingtonVA.USA1990.8.C.L.CrofcheckF.A.PayneandM.PinarMengucCharacterizationofmilkpropertieswitharadiativetransfermodelAppl.Opt.4110102002.9.P.WalstraTurbidimetricdeterminationofthefatcon-tentofmilkNeth.MilkDairyJ.192661965.10...and..2006inRussian.11.A.V.KalininV.N.KrasheninnikovandA.V.PotapovStudyofdispersedsystemsbyNIRspectroscopyandPLSregressiontechniqueproteinfractionsinmilkandreversedmicellessolutionsChemometr.Intell.Lab.Syst.97332009.doi10.1016j.chemolab.2008.08.00712.A.KalininV.KrasheninnikovS.SadovskiyH.DenisovichH.YurovaandV.KrlvtsunCalibrationmodelsformulti-componentquantitativeanalysesofdairywiththeuseoftwodifferenttypesofportablenearinfraredspectrometerJ.NearInfraredSpectrosc.1632008.doi10.1255jnirs.79713.R.TsenkovaS.AtanassovaK.ToyodaY.OzakiK.ItohandT.FearnNear-infraredspectroscopyfordairymanagementmeasurementofunhomogenizedmilkcompositionJ.DairySci.821123441999.doi10.3168jds.S0022-03029975484-614.I.V.YaroslavskyA.N.YaroslavskyT.GoldbachandH.-J.SchwarzmaierInversehybridtechniquefordetermin-ingtheopticalpropertiesofturbidmediafromintegrat-ing-spheremeasurementsAppl.Opt.3534341996.15.Ph.LavenAcomputerprogramforscatteringoflightfromasphereusingMietheoryandtheDebyeseries2011httpwww.philiplaven.commieplot.htm.Accessed4April2011.16...and..-1095641320112011..292011inRussian.17.S.WoldM.SjstrmandL.ErikssonPLS-regressionabasictoolofchemometricsChemometr.Intell.Lab.Syst.582001109.doi10.1016S0169-74390100155-118.R.G.WhiteldM.E.GergerandR.L.SharpNear-infraredspectrumqualicationviaMahalanobisdis-tancedeterminationAppl.Spectrosc.41719871204.doi10.1366000370287444757219.12.N88- A.V.KalininV.N.KrasheninnikovandV.M.KrivtsunJ.NearInfraredSpectrosc.2135412013419-hpec.europa.eufoodinternaonaltradedocsmilk_dairy01_ru.pdfAssessed13thFebr20132008inRussian.20.A.BogomolovS.DietrichB.BoldriniandR.KesslerQuantitativedeterminationoffatandtotalproteininmilkbasedonvisiblelightscatterFoodChem.1342012412.doi10.1016j.foodchem.2012.02.077 JOURNALOFNEARINFRAREDSPECTROSCOPY43ISSN0967-0335IMPublicationsLLP2013doi10.1255jnirs.1038AllrightsreservedMacroporousresinshavetheadvantagesofrelativelylowcosthighadsorptioncapacityeasyregenerationetc.makingthempopularforpharmaceuticalapplications.Theadsorptionelutionprocessonmacroporousresinsisoneoftheefcientenrichmentmethodswithamoderatepurificationeffect.Ithasbeenusedfortherecoveryandconcentrationofsmallmoleculesincludingpurificationofglycosidesflavonoidsetc..13Theadsorptioneffectofmacroporousresinisdeterminedbysurfaceadsorptionandsieveclassicationhydrogenbondinginteractionssurfaceelectricalpropertyetc.4Whentheadsorptionreachesleakpoint5theadsorptionaffinitydecreasesandthesolutesbegintoleakfromtheresininlargeamounts.Soitiscriticaltocontroltheleakpointinordertoensureadequateadsorp-tionandreducewasteofthesamplesolution.Meanwhiletheapplicationofmulti-columnsystemswithmacroporousRapidandquantitativedetectionmethodforacteosideduringchromatographicpurificationofadhesiverehmannialeafextractusingnearinfraredspectroscopyandchemometricsYeJinaXuesongLiuaLianjunLuanaGuangmingQingbYingZhongbandYongjiangWuaaCollegeofPharmaceuticalSciencesZhejiangUniversityHangzhou310058China.E-mailyjwuzju.edu.cnbSichuanMedicoPharmaceuticalCompanyShifang618400ChinaThefeasibilityofrapidandnon-destructiveanalysisofacteosidethemainactivecomponentofadhesiverehmannialeafinthemacroporousresinadsorptionandelutionprocessbyFouriertransformnearinfraredspectroscopytogetherwithdifferentchemometricsmethodswasinvestigatedinthisstudy.Theusablespectralregion55006200cm1wasidentiedspectralpreprocessingincludingSavitzkyGolaysmoothingderivativewasemployedandspectraldimensionwasalsoreducedthroughprincipalcomponentanalysisforarticialneuralnetworksANNandleastsquaresupportvectormachinesLS-SVMmethods.Themultivariatecalibrationmodelsbasedonprincipalcomponentregressionpartialleast-squaresregressionANNandparticleswarmoptimisation-basedLS-SVMweredevelopedtocorrelatethepretreatedspectraldataandthecorrespondingacteosideconcentrationsdeterminedbyhighperformanceliquidchromatography.Forallregressionmodelsthecoefcientsofdeterminationforalldatasetscalibrationvalidationandpredictionsetwerehigherthan0.988indicatingthatthefourmethodswereeffectiveforbuildingacteosidecalibrationmodelsusingsamplestakenfromeithertheadsorptionortheelutionprocess.HoweveraccordingtotherelativestandarderrorofpredictionvaluesforthepredictionsetsamplestheLS-SVMmodelwasslightlymoreaccuratethanthemodelsobtainedusingtheotherregressiontechniques.KeywordsnearinfraredspectroscopypartialleastsquaresregressionarticialneuralnetworksleastsquaresupportvectormachinesmacroporousresinIntroductionY.Jinetal.J.NearInfraredSpectrosc.2143532013Received3September2012Revised19November2012Accepted29November2012Publication13February2013 44RapidandQuantitativeDetectionofActeosideDuringPuricationofAdhesiveRehmanniaLeafExtractresinsinapharmaceuticalindustrymakesthedetermina-tionofsaturationpointimportant.Inadditiontheelutionendpointalsoneedstobecontrolledtosaveelutionsolventandmakesuretheelutionprocessiscompleted.TotalglycosidesfromadhesiverehmannialeafARLhaveadirecteffectontheglomerularblockingtheformationofimmunecomplexesinsitusuppressionresistingglomerularcapillarybasementmembraneinjuryinhibitingglomer-ularmesangialcellproliferationreducingrenaldialysisandimprovingthequalityoflife.6ActeosideamainactivephenylethanoidglycosideintotalglycosidesofARLexhibitsantioxidativeanti-inflammatoryandneuronalprotectiveactivitiesbothinvitroandinvivo.7TotalglycosidesofARLareproducedbysolventextractionandpuricationusingmacroporousresincolumnchromatography.Puricationisthekeyproductionprocessforitsquality.Inthepurifica-tionprocessconstantdetectionofthetotalglycosidesoracteosideconcentrationisrequiredforqualitycontrol.Themostcommonmethodusedtodetecttheleakpointinapharmaceuticalindustryinvolvesaferricchloridecolourreaction.Howeverthecolourreactionissusceptibletointerferenceandcannotbeusedforquantitativeanalysis.CurrentlyquanticationofacteosidereliesheavilyonhighperformanceliquidchromatographyHPLC.Althoughreli-ableandrelativelyaccurateHPLCistime-consumingandrequirescomplexsamplepreparationlimitingitsapplicationtorapiddetectionandprocesscontrol.Thereforeafastandaccuratemethodisrequiredtospeedupthedetermina-tionoftheactiveglycosidetopermitqualitycontrolduringpharmaceuticalproduction.NearinfraredNIRspectroscopyisasimplefastandnon-destructivetechniquethatenablestheanalysisofsampleswithoutcomplicatedpretreatmentswhichresultsinsubstantiallydecreasedanalysistimerelativetotraditionalanalyticalmethodssuchaschromatographictechniques.NIRspectroscopyismainlyusedtorecordinformationintheovertoneandcombinationbandregionsofthespec-trumwhicharisefromfundamentalvibrationsinthemid-infraredregion.8DirectquanticationanalysisbasedonthecomplexityandhighdimensionofNIRspectraldataaredif-cult.InthisstudywecomparedthechemometricsmethodsprincipalcomponentregressionPCRpartialleast-squaresPLSregressionarticialneuralnetworksANNandleastsquaresupportvectormachinesLS-SVMtoextractspec-tralfeaturesandinvestigatecorrelationsbetweenthespec-traldataandcomponentconcentrations.PCRandPLSaresuitableforlinearanalysiswhileANNandLS-SVMhaveadvantagesfornon-linearanalysis.Recentlyanumberofcomparativestudiesonthesetechniqueshavebeeninvesti-gatedbasedonvariousdatasets.Kovalenkoetal.9comparedPLSLS-SVMandANNmethodsfortheestimationofaminoacidcompositioninsoybeansandconcludedthattheperfor-manceofPLSandLS-SVMwassignicantlybetterthanthatofANN.Sunetal.10comparedthechemometricsmethodsPCRPLSANNandLS-SVMtoassesstheinternalqualityofpearsbymeasuringsolublesolidcontentSSCandaciditypHon-line.TheresultsshowedthatPLSandLS-SVMweresuperiortoothermethodsforassessinginternalqualityofpears.BalabinandSaeva11comparedtheperformanceoflinearmultiplelinearregressionMLRPCRandPLSandnon-linearcalibrationtechniquespolynomialandsplinePLSandANNforpredictionofbiodieselpropertiesfromNIRspectra.TheyfoundthattheANNapproachwassuperiortothelinearandquasi-non-linearpolynomialandsplinePLScalibrationmethods.Accordingtotheresultstheyconrmedthatbiodieselwasahighlynon-linearobject.InanotherpaperBalabinandLomakinar12reportedthatthesupportvectormachineSVM-basedapproacheswerecomparabletoANNinaccuracy.Howevertheyweremuchmorerobustandwererecommendedforpracticalindustrialapplica-tions.Pengetal.13proposedahybridmultivariatecalibrationalgorithmwhichwasthecombinationofprincipalcompo-nentanalysisPCAparticleswarmoptimisationPSOandLS-SVMPCA-PSO-LS-SVM.InthisalgorithmPCAwasemployedtoreducethedimensionofrawspectraandPSOwasusedtooptimisetheparametersofLS-SVMmodelinthecalibrationset.TheresultsshowedthatcomparedwiththePLSRalgorithmthePCA-PSO-LS-SVMalgorithmcouldgreatlyimprovethepredictionprecisionofamodelrevealingthatitwasanefcienttoolforNIRspectraregression.Guoetal.14alsoreportedthatPSO-basedLS-SVMwithradialbasisfunctionRBFkernelwassuperiortoconventionalmethodsincludingANNandPLSRmodels.Accordingtothepapersmentionedaboveonecanassumethatthedifferentconclusionsobtainedfromdifferentstudiesresultedfromthedifferencesinthenatureofthenon-linearities.15TheaimofthisstudywastoapplyFouriertransformnearinfraredFT-NIRspectroscopictechniqueswithHPLCasareferencemethodtoquantitativeanalysisofacteosideduringchromatographicpurication.ThefeasibilityofusingNIRtodetermineacteosideinthemacroporousresinadsorptionandelutionprocesseswasalsoexplored.TheperformanceoflinearPCRandPLSandnon-linearcalibrationtechniquesANNandLS-SVMwerealsogenerallycompared.MaterialsandmethodsMaterialsARLcrudematerialswereprovidedbytheSichuanMedicoPharmaceuticalCompanySichuanChina.Acteoside99.5puritywaspurchasedfromtheNationalInstitutefortheControlofPharmaceuticalandBiologicalProductsBeijingChina.D101macroporousresinwassuppliedbyHuipuChemicalApparatusCo.LtdHangzhouChina.Ethanol95vvusedastheextractionandelutionsolventwasofmedicalgrade.MethanolchromatographicallypuregradewaspurchasedfromMerckDarmstadtGermanyforHPLCanalysis.WaterwaspuriedbyMilliporewaterpuricationdeviceMilliporeCorp.BillericaMAUSA.Allotherreagentswereofanalyticalgrade. Y.Jinetal.J.NearInfraredSpectrosc.214353201345PretreatmentofmacroporousresinTheD101macroporousresinwaspretreatedaccordingtothemanufacturersinstructionstoremovethemonomersandporogenicagentstrappedinsidetheporesduringthesynthesisprocess.Theadsorbentbeadsweresoakedwith10timesvolumeof95ethanolvvfor24handthenwashedwith95ethanoluntilnowhiteturbidsolutionwasobserved.Priortousethechromatographiccolumnwaswashedwithdeionisedwateruntiltheethanolwasthoroughlyreplacedwithdeionisedwater.16PreparationofcrudeARLextractARL2.8kgwasextractedwith28Lof80aqueousethanolcontaining0.5aceticacidvvtwice90mineachtimeinamulti-functionextractor40L.Theextractionsolutionwasconcentratedbyarotaryevaporatorat50C.Theconcentratedsolutionwascentrifugedat6000rpmfor20min.Thesuperna-tantwasthenaddedto3.3Lwaterforadsorptionandelutionexperiments.DynamicadsorptionandelutionprocessesDynamicadsorptionandelutionprocesseswerecarriedoutinaglasscolumni.d.2cm30cmwet-packedwith60mLofD101macroporousresin.ThecrudeARLextractwasloadedontothecolumn.Theleakowratewassetat4BVh14mLmin1BVstandingforbedvolume.Afterfeedingfor12mintheleaksolutionswerecollectedevery25minandthevolumeofeachsamplewas4mL.Theloadingwasstoppedwhenthefeedsolu-tionreachedapproximately6BV.Theadsorbate-ladencolumnwasrstwashedbyde-ionisedwatertoremovetheresidualsolutionandcomponentswithstrongpolarityandthenelutedwith50aqueousethanolsolutioncontaining0.5aceticacidvv4BVattheowrateof4BVh1.Afterelutingfor4mintheefuentswerecollectedevery25minandthevolumeofeachsamplewas4mL.Allthedynamicexperimentswereperformedatroomtemperature.High-performanceliquidchromatographyAvalidatedHPLCmethod17withaslightmodificationwasappliedtothequantitativedeterminationofacteosideinsamplesandtheresultswereusedasreferencedataforNIRanalysis.ChromatographicanalysiswasperformedonaWaters2695SeparationModuleMilfordMAUSA.ActeosidewasseparatedonaDiamonsilC18column250mm4.6mmi.d.particlesize5mDikmausingmethanol1phosphoricacid3565vvasthemobilephaseataflowrateof1mLmin1.Theinjectionvolumewas20Landalldetermi-nationswereperformedat40C.TheultravioletUVdetec-tionwavelengthwasat334nm.Theacteosideconcentrationwascalculatedusingacalibrationcurveobtainedafterlinearregressionofthepeakareavsacteosideconcentration.Thecalibrationcurvewaslinearintherangeof0.028mgmL1and0.45mgmL1withthecoefcientofdeterminationof0.9998.SpectralmeasurementNearinfraredtransmittancespectraofsamplesfromtheadsorptionandelutionprocesseswerecollectedat16cm1intervaloverthespectralrangeof400010000cm1usinganAntarisThermoFisherScienticInc.MadisonWIUSAFT-NIRspectrometer.Eachsamplewasscannedwitha2mmpathlengthandaspectrumobtainedbyaveraging32scans.Allspectrawerecollectedinabsorbancemode.Theaveragespectrumfromeachtriplicatemeasurementasthenalspec-trumofeachsamplewasusedtoquantifyacteosideconcen-tration.DataacquisitionspectralpretreatmentsandPCAwereperformedusingtheRESULTsoftwaresuiteversion3.0ThermoNicoletUSATQanalystversion8.0ThermoNicoletUSAandMatlabversion7.5theMathWorksInc.NatickMAUSA.SamplesInthisstudywerepeatedtheadsorptionelutionexperi-mentsfourtimesandobtained84samplesleaksolutionsfromtheadsorptionprocesses82sampleseffluentsfromtheelutionprocesses.ThesesampleswereusedforNIRdatacollectionandHPLCanalysis.Samplesfromonerandomisedexperimentwithnoextremeacteosidevalueswereselectedintoavalidationsetandsamplesfromtheotherthreeexperimentswereusedasthecalibrationset.18AnewadsorptionelutionexperimentwasperformedafterthecalibrationandvalidationoftheNIRmodels.Samplesfromthenewexperimentformedapredictionsamplesetandwereusedtoassessthepredictiveabilityofmodels.ThestatisticalresultsofcalibrationvalidationandpredictionsetarelistedinTable1.AdsorptionElutionCalibrationsetValidationsetPredictionsetCalibrationsetValidationsetPredictionsetSamplenumber632121612121MeanmgmL18.3598.6768.6454.6984.7954.574RSD63.5957.4861.15121.3117.7125.8MaxmgmL115.17814.41414.45417.27816.07816.843MinmgmL10.00840.03140.02290.01490.11130.0454Table1.Statisticalresultsofcalibrationvalidationandpredictionsetsamples. 46RapidandQuantitativeDetectionofActeosideDuringPuricationofAdhesiveRehmanniaLeafExtractChemometricsmethodsInthisstudywemadeageneralcomparisonoftheaccuracyandrobustnessoffourmultivariatecalibrationmodelsPCRPLSRANNandLS-SVM.TheperformanceoftheestablishedmodelswasassessedintermsofrootmeansquareerrorofcalibrationRMSECcoefcientofdeterminationforthecalibrationsetR2coefcientofdeterminationforthevalida-tionandpredictionsetr2rootmeansquareerrorofpredic-tionRMSEPrelativestandarderrorofpredictionRSEPandresidualpredictivedeviationRPD.1920ThecalibrationmodelswhichhadthehighestR2orr2thelowestRMSECandRMSEPalsotheleastdifferencebetweenRMSECandRMSEPwereconsideredoptimal.TheRPDdenedasthestandarddevia-tionSDofthereferencevaluesdividedbythestandarderrorofperformanceSEPisanotherimportantindexusedtochecktheperformanceofamodel.RelativelyhighRPDvaluesindicatemodelshavinggreaterpowertopredictthechem-icalcomposition.21RSEPwasalsocalculatedforthecalibra-tionvalidationandpredictionsettoassessthequalityoftheresults.22TheRSEPiscalculatedasbelow-22100iiiCCRSEPCwhereCiistheacteosideconcentrationmeasuredbyHPLCandCiistheacteosideconcentrationpredictedbyNIR.PCRPCRisbasedonPCA.PCAisaneffectivedataminingtechniquewhichextractsmaininformationfromNIRspectraandcompressesintoafewprincipalcomponentsPCsreducingthenumberofvariables.ThereforethePCAeigenvectorsandscoreswhichrepresentthelargestcommonvariationsamongallthespectrainthecalibrationdataarecalculatedrst.ThentheselectedPCAscoresareregressedagainsttheconstituentconcentrationsusingaregressionmethod.23PLSRPLSRwhichisbasedonmultivariateregressionisthemostcommonlyandwidelyusedlinearregressionmethodforitssimplicityspeedrelativelygoodperformanceandeasyaccessibility.24TheprinciplebehindthePLSalgorithmistoextracttheimportantinformationfromvariationofbothopticalXandreferencedataYandcompressitinasetofnewindependentlatentvariablesLVsequaltoPCs.ThedetailedinformationforPLScanbefoundintheliterature.2526DuetoitssimplicityandsmallvolumeofcalculationsthePLSapproachisoneofthemostpopularchemometricsalgorithmsforspectroscopicdataanalysis.27ThePLSmethoditselfaswellasPCRisalinearmethodofdataanalysis.Howeveroverlapamongsignalsandviola-tionsoftheBeerLambertlawoftencausesnon-linearities.Othersourcesofnon-linearityinNIRspectralmeasure-mentsarenon-linearinstrumentresponsesandinteractionsbetweencomponents.28Thenon-linearitiesofNIRspectramaymakePLSmodelspoorintermsofpredictiveability.SincethecompositionoftraditionalChinesemedicineTCMisquitecomplexthedataobtainedbyNIRinstrumentscouldcontainnon-linearities.Thereforenon-linearmethodscouldbesuperiortolinearcalibrationtechniques.Non-linearrelationscanbemodelledbyPLSinalimitedwaybyconsideringmorelatentvariables.IncludingmoreLVsinthemodelmayallowabetterfittothetrainingsetsamplesbutthepredictionoftheothersamplesmaybecomeworse.Thisphenomenoniscalledover-ttingofamodel.HoweverittoofewLVsareusedinthemodelunder-ttingwilloccur.ExactlythesamecanbesaidaboutthePCRtech-nique.TotesttherobustnessofthePCRandPLSmodelsaleave-one-outcrossvalidationwasusedandtheoptimumnumberofLVsincludedinthemodelswasdeterminedbyPRESSpredictedresidualerrorsumsquare.ComputationsofPCRandPLSRwereperformedusingTQanalystsoftwareversion8.0ThermoNicoletUSA.ANNANNiswidelyusedformodellingthenon-linearbehaviourofaprocessbecauseitallowsagreatdealofflexibilityindeterminingmodelstructures.Aneuralnetworkisafunctiondescribedbythenumberofhiddenlayersthenumberofinterconnectedneuronsateachlayerwiththeirtransferfunctionsandasetofweightsincludingbiastermsassignedtolinksconnectingtheneurons.VarioustypesofANNhavebeenproposedthatdifferintheirstructuresandlearningalgorithms.29Amongthesetypesfeed-forwardback-propagationnetworkBP-ANNisthemostextensivelyappliedmatureANNinpredictionanditstrainingalgorithmisthewell-knowngradientdescentmethodGDM.Inthisstudyatwo-layerBPnetworkwasemployedwhichcouldproducesolutionsarbitrarilyclosetotheoptimalsolution.ThemainlimitingfactoroftheANNapproachisaccesstoasufcientnumberoftrainingsamples.TheavailablesamplesetshouldberelativelylargeforeffectiveANNtraining.WhenANNisutilisedwithNIRspectraldatawherethenumberofinputvariableswavenumbersisusuallylargeandthenumberoftrainingsamplesislimiteditispracticaltoreducetheinputdimensionality.InthisstudythedimensionalityoftheNIRspectrawasreducedbyPCA.SincenetworksaresensitivetothenumberofinputneuronsequaltoPCsandneuronsintheirhiddenlayersseveraltrainingofnetworkswasperformedwithdifferentnumbersofinputneuronsfrom1to15andhiddenlayerneuronsfrom1to30.TheperformanceofeachANNmodelwasevaluatedbycalculatingthemeansquareerrorMSEtheaveragesquarederrorbetweenthenetworkoutputsandthetargetoutputs.ThecommonlyusedtransferfunctionsforBPnetworksarelogsiglogsigmoidtransferfunctiontansighyperbolictangentsigmoidtransferfunctionandpurelinlineartransferfunction.ThetansigtransferfunctionwhichcanproducebothpositiveandnegativevaluestendedtoyieldfastertrainingandsmallerMSEsthanthelogsigtransferfunctionwhichcanproduceonlypositivevalues.Inthisstudythetansigandpurelinfunctionswereemployed. Y.Jinetal.J.NearInfraredSpectrosc.214353201347OncetheBPnetworkwasconstructedandtheweightsandbiaseswereinitialisedthenetworkwasreadyfortraining.NeuralNetworkToolboxforMATLABoffersseveraltrainingalgorithmssuchastraingdtraingdxtraingdatrainrptrainlmetc.whichareusedfortrainingBPnetworks.ThetrainlmalgorithmupdatesweightandbiasvaluesaccordingtoLevenbergMarquardtoptimisationwhichisoftenthefastestback-propagationalgorithmwithitslargestmemoryrequire-ment.Accordingtothesatisfyingstatisticalresultsthetrainlmalgorithmwasselectedfortrainingtheacteosideconcentra-tionANNmodel.Theothertrainingparameterswereselectedasdefault.ThedetailedproceduresforANNtrainingcanbefoundintheliterature.9121430LS-SVMLS-SVMwasintroducedbySuykensetal.31asreformulationstostandardSVM3233whichleadstosolvinglinearKarushKuhnTuckerKKTsystemsforclassicationandregressionproblems.LS-SVMsimpliesthetrainingprocessofstandardSVMtoagreatextentbyreplacingtheinequalityconstraintswithequalitycounterpartsandaleastsquareslinearsystemasalossfunctionisadoptedinsteadofaquadraticprograminstandardSVM.34ForLS-SVMtheparametersintheregularisationitemandkernelfunctionso-calledhyper-parametersplayacrucialroleinthealgorithmperformance.Overthelastyearsmanyapproachesforturningthehyper-parame-tershavebeenproposedintheliterature.3540PSOwhichwasdevelopedbyKennedyandEberhartin199541isastochasticglobaloptimisationtechniqueinspiredbythesocialbehaviourofflockingbirds.ThePSOpossessesthecapabilitytoescapefromlocaloptimaiseasytobeimplementedandhasfewerparameterstobeadjusted.InPSOaswarmconsistsofindividualscalledparticleswhichrepresentsapotentialsolutiontotheproblem.Eachparticleiesaroundinamultidimensionalsearchingspacewithavelocitydirectingtheyingandadjustsitspositionaccordingtoitsownbestpreviousexperiencepbestandtheexperienceofallothermembersgbest.41Alltheparti-cleshavetnessvalueswhichareevaluatedbytheprede-nedtnessfunctions.Accordingtotnessinformationthepopulationisupdatedandtheparticlesmovetowardsthebettersolutionareaswhilestillhavingtheabilitytosearchawideareaaroundthebettersolutionareas.ThePSOalgo-rithmhasbeenfoundtobefastandrobustinsolvingnon-linearnon-differentiableandmultimodalproblems.42Thein-depthmathematicaldescriptionandexecutivestepsofthePSOcanbefoundinLinetal.43AlthoughtheNIRdataofthespectralregionofinterestcouldbeappliedasinputfortheLS-SVMmodelthetrainingtimeincreaseswiththesquareofthenumberoftrainingsamplesandlinearlywiththenumberofvariables.27InordertoimprovethetrainingspeedandreducethetrainingerrorscoresoftherstseveralPCsobtainedfromPCAwereappliedasinputofLS-SVMmodels.InthesamewaywithANNthenumberofPCswasselectedonthebasisoflowestRMSECRMSEPandhighestR2intheoutput.RBFwasusedasakernelfunctionforLS-SVMmodelbuildingwhichcanhandlethenon-linearrelationshipsbetweenthespectraandacteo-sideconcentrationandbeabletoreducethecomputationalcomplexityofthetrainingprocedure.ThequalityofLS-SVMforregressiondependsongammaandsig2parameters.Theregularisationparametergammadeterminesthetrade-offbetweenminimisingthetrainingerrorandminimisingmodelcomplexity.Theparametersig2isthebandwidthofRBFandimplicitlydenesthenon-linearmappingfrominputspacetosomehighdimensionalfeaturespace.Inthisstudythroughinitialexperimentsgivenaswarmof100particleseachparticleii12100isassoci-atedwithapositionvectorxixi1xidxiDwhichisafeasiblesolutionfortheoptimalproblem.DisthenumberofdecisionparametersoftheoptimalprobleminthisstudyD2gammaandsig2.Thevelocityfortheithparticleisrepresentedasvivi1vidviD.MSEissettobethetnessfunctionwhichvarieswiththeLS-SVMparameters.InPSOthevelocityoftheparticleisacceleratedinthedirec-tionsoftheselocationsofgreatesttnessaccordingtothefollowingequationvidwvidc1r1pibestdxidc2r2gibestdxidwherevidisthevelocityoftheparticlexidistheparticlescoordinateintheDthdimensionpibestdisthelocationinparam-eterspaceofthebesttnessreturnedforaspecicagentgibestdisthelocationinparameterspaceofthebesttnessreturnedforanentireagentc1andc2arelearningfactorsandchosentobec1c22r1andr2aretworandomparametersbetween0and1andwistheinertiaweightchosenheretobe0.5.AfterimplementingPSOtheoptimalhyper-parametersgammaandsig2fortheLS-SVMalgorithmcouldbeobtained.TheLS-SVMregressionmodelwasestablishedusingtheLS-SVMlab1.5MATLABtoolbox.44ResultsanddiscussionHPLCanalysisTheleakpointwaspresumablyreachedwhentheacteo-sideconcentrationinleaksolutionwas10oftheoriginalconcentration.5Continuouslyloadingtheextractledtorapiddecreaseofadsorptioncapacityandtheadsorptionkineticreachedasteadyplateau.Thesaturationpointwaspresum-ablyreachedwhentheacteosideconcentrationinleaksolu-tionequalledtheoriginalconcentration.ItcanbeseenfromFigure1athattheleakpointwasachievedwhenthefeedsolutionreached11.2BVandtheadsorptionofacteosidereachedasteadyplateauwhenthefeedsolutionwasapprox-imately5.5BVsaturationpoint.Furthermoreacteosidewasalmostelutedoutofthecolumninabout3.5BVwhentheacteosideconcentrationinefuentsroughlyequalledtozeroduringthenalstageelutionendpointasshowninFigure1b. 48RapidandQuantitativeDetectionofActeosideDuringPuricationofAdhesiveRehmanniaLeafExtractSpectralanalysisFigure2showstheNIRspectraoftheleaksolutions105samplesandtheefuents103sampleswhichillustratesthelowestmolecularabsorptivitiesinthehighwavenumberregion720010000cm1withhighervaluesintherstover-toneregion55006200cm1andstillhigherabsorbancelevelsinthecombinationregion40004800cm1.Thespectraofsamplesfromtheadsorptionprocesseswerelessvariablethanspectrathatfromtheelutionprocessesowingtothefactthatthecompositionsincludingsolventintheefuentsvariedsignicantlyduringtheelutionprocesses.TheOHandtheCHbondsareabundantinthemolecularstructureofacteosideseeFigure1sothattheNIRspectracanbeconsideredtoreflectthechemicalinformationofacteosidesufficiently.HowevertheNIRspectraexhibitintenseabsorptionbandsat6944cm1relatedtotherstOHovertoneandat5155cm1relatedtothecombinationofstretchinganddeformationoftheOHgroupinwater.18Thesetwowaterregions50005500cm1and62007400cm145whicharetypicaloftheNIRspectrumofanaqueoussolutionmakethetechniquedifculttodetermineacteosidesincetheysaturatethedetectorandmaskanyotherbandspresentinthesespectralranges.The40005000cm1regioncorre-spondingtocombinationsoffingerprintabsorptionswithCHOHandNHstretchingmodes46werenotemployedinspectralanalysisduetothezerotransmissivityandthefactthattheyweresaturated.47Besidesthe740010000cm1regionassignedtosecondandthirdovertonesischar-acterisedbylowintensityandlowsignal-to-noiseratio.48Thereforeaccordingtothereasonsmentionedabovetheregionof40005000cm150005500cm162007400cm1and740010000cm1wereexcludedduringquantitativeacteosideanalysis.Thespectralregionactuallyemployedtoestablishquantitativemodelswas55006200cm1seeFigure2whichencompassesbandsoriginatingfromtherstovertonesoftheCHstretchingmode.49Althoughthisspectralregionisweakandoralwaysconcealedbytheintenseabsorptionbandsofwateritcontributessubstan-tiallyenoughtoformthebasisforquantitativeacteosideanalysis.5051Asdetectionofthesamecomponentacteo-sidethesameband55006200cm1wasselectedforbothadsorptionandelutionprocessesinthisstudy.Toconrmthecorrectnessofthespectralregionselectionthecorrelationcoefcientsofthespectrawereinvestigated.Thecorrelo-gramisshowninFigure3inwhichitcanbeseenthatthevariableswithhighercoefcients0.8aremostlydistrib-utedintheselectedregion.DerivativesarepracticalintheNIRregionforreducingpeakover-lapeliminatingconstantandlinearbaselinedriftandrevealingpeaksinoriginalspectra.TherstderivativespectraoftheadsorptionandelutionsamplesclearlyrevealedanumberofabsorptionsthatmaybeascribedtotheacteosideasshowninFigure4.Toavoidenhancingthenoiseallderiva-tivespectraweresmoothedusinga21-pointSavitzkyGolaylter.52Figure2.NIRspectraofatheleaksolutionsfromtheadsorptionprocessesandbtheefuentsfromtheelutionprocesses.Figure3.CorrelogramoftheNIRspectraandtheacteosideconcentrationsaadsorptionprocessandbelutionprocess.Figure1.aAdsorptionandbelutionkineticscurvesforacteosideonD101resinandthechemicalstructureofacteoside.ab Y.Jinetal.J.NearInfraredSpectrosc.214353201349ComparisonofregressionmethodsFortheANNmodeloftheadsorptionprocesstherstthreePCswereselectedasinputofnetworksandtheseexplained99.13ofthetotalvariance.MeanwhiletherstsevenPCswhichcouldexplain99.61ofthevariationweresetasinputoftheANNmodelofelutionprocess.AccordingtocertainstatisticalparameterssuchasRMSECRMSEPandR2etc.theoptimaltopologyfortheANNmodelofadsorptionprocesswasfoundtobe3201i.e.threeinputneuronsasinglehiddenlayerwith20neuronsandoneoutputneuron.TheoptimaltopologyfortheANNmodelofelutionprocesswasfoundtobe7201.ForPSObasedLS-SVMtheoptimalvaluesofgammaandsig2werecalculatedasfollowsfortheLS-SVMmodelofadsorptionprocessgamma1.3552104sig21.1032105fortheLS-SVMmodelofelutionprocessgamma5.1149106sig26.8842103.UsingtheoptimumparametersforPCRPLSRANNandLS-SVMthemultivariatecalibrationmodelsforacteosidewereestablished.ThestatisticsoftheestablishedmodelsofadsorptionandelutionprocessesarelistedinTable2.HPLCmeasuredvaluesvsNIRpredictedvaluesforacteosideconcentrationincalibrationandvalidationsetsareprovidedinFigure5.BasedondatafromTable2onecanseethatforallmethodsR2andRPDofthecalibrationsetwerehigherthan0.988and9.229respectivelyandtheRMSECandRSEPwerelessthan0.429and8.332respectively.Forthevalidationsetr2andRPDwerehigherthan0.992and10.505respec-tivelyandtheRMSEPandRSEParelessthan0.438and7.18respectively.ThedifferencesamongPCRPLSRBP-ANNandLS-SVMresultswereverysmallrevealingthateachofthefourmethodswaseffectiveforbuildingacteosidecalibrationmodelsforsamplesfromboththeadsorptionandtheelutionprocesses.ComparedwiththemodelsoftheadsorptionprocessthecalibrationmodelsfortheelutionprocessrequiredmorePCs.Thegeneraltrendisthatthemorecomplicatedthequalitythatisthegreaterthenon-linearitythegreaterthenumberofPCsneededtoextractallnecessaryinformationandtotakeintoaccountthedeviationfromthelinearspectrumpropertydependence.Thereforetheeffluentswererelativelyhighnon-linearobjects.ItshouldbenotedthatfortheadsorptionprocesstheoptimumnumberofPCsusedinthePCRmodelwas25seeTable2indicatingthatPCRasalinearcalibra-tionmethodcanhandlenon-linearitybyaddingmorePCstothemodel.WithrespecttoRSEPandRPDvaluesthequantitativemethodscanbearrangedinthefollowingorderforthecali-brationsetBP-ANNLS-SVMPLSRPCRforthevalida-tionsetPLSRLS-SVMPCRBP-ANN.WhileBP-ANNhadthebestcalibrationresultsthepredictionresultswereslightlyworseandclosertotheresultwhenusingPCR.ThisobservationcanbeexplainedbythefactthattheANNmethodhasthetendencytoover-tandthisbehaviourcanlowerthegeneralisationabilityofthenetwork.ThereforeamongtheseregressionmethodologiesPLSRandPSO-basedLS-SVMweresufcientlyaccurateandrobusttobeusedforacteosideanalysisandarerecommendedforpracticalimplementation.ThepredictiveabilityoftheadsorptionandelutionmodelswasevaluatedusingthepredictionsetTable3.Theresultsweresatisfactoryandforthefourquantitativemethodsr2fortheadsorptionandelutionprocesswerehigherthan0.992and0.996respectively.AmongthesemodelstheFigure4.Firstderivativespectraofatheleaksolutionsfromtheadsorptionprocessesandbtheefuentsfromtheelutionprocessesandthespectralrangeforbuildingmodels.CalibrationsetValidationsetNumberofPCsRMSECR2RSEPRPDRMSEPr2RSEPRPDPCRAdsorption50.3610.9953.65814.5960.4240.9954.23311.485Elution250.4120.9888.3329.2290.4380.9927.1810.505PLSRAdsorption60.2350.9982.37422.4760.2230.9982.2421.845Elution70.4290.9945.84313.160.4330.9955.92712.726BP-ANNAdsorption30.1590.9991.61133.1170.4240.9954.26211.478Elution70.0930.9991.26960.5770.2940.9966.85810.997LS-SVMPSObasedAdsorption50.1810.9991.83629.0560.2030.9982.04823.893Elution70.4030.9955.48414.0210.3960.9985.46914.178Table2.Statisticsoftheestablishedmodelsofadsorptionandelutionprocesses. 50RapidandQuantitativeDetectionofActeosideDuringPuricationofAdhesiveRehmanniaLeafExtractFigure5.HPLCmeasuredvaluesvs.NIRpredictedvaluesforacteosidecalibrationsetvalidationset.Corr.Coeffandslopevaluesareforthecalibrationandvalidationsamples.PredictionsetRMSEPr2RSEPRPDPCRAdsorption0.6220.9986.1548.293Elution0.3840.9965.30514.614PLSRAdsorption0.4850.9954.82110.114Elution0.3390.9974.6816.568BP-ANNAdsorption0.4960.9924.92610.403Elution0.4740.9976.55111.835LS-SVMPSObasedAdsorption0.3540.9963.51314.587Elution0.2520.9983.45421.837Table3.Statisticsofthepredictionsetofthefourquantitativemodels. Y.Jinetal.J.NearInfraredSpectrosc.214353201351LS-SVMmethodprovidedthebestpredictiveability.ThetrendsofacteosideconcentrationpredictedbyNIRandmeasuredbyHPLCusingLS-SVMwerealmostperfectlysuperimposedasshowninFigure6.ItcanbeseenfromFigure6athatthepredictedleakpointwasachievedwhenthefeedsolutionreached1BVandtheadsorptionofacteo-sidereachedasteadyplateauwhenthefeedsolutionwasapproximately5.2BVdemonstratingthatthesaturationpointwasachieved.MoreoverFigure6bshowsthatalltheacteosidehadbeenelutedfromthecolumninabout3.5BVelutionendpoint.ConclusionsInourworkPCRPLSRBP-ANNandPSObasedLS-SVMasfourlinearandnon-linearregressionmethodswereinvesti-gatedtobuildquantitativemodelsforpredictionofacteosideduringadsorptionandelutionprocesses.Allofthesefourcalibrationmodelswerecomparabletoeachotherinaccuracyandprovedsuitableforpurpose.HoweveraccordingtothepredictiveresultsofthepredictionsetthePSObasedLS-SVMapproachhadhigheraccuracy.InsummarythisstudydemonstratedthatNIRspectroscopyisapromisingtechniqueforquanticationofacteosideinthemacroporousresinadsorptionandelutionprocessesofARL.WebelievethattheresultspresentedhereinwillhelpfuturechemometricinvestigationsandresearchesinNIRspectro-scopyofmulti-componentsystems.AcknowledgementTheauthorsthanktheSichuanMedicoPharmaceuticalCompanyforprovidingadhesiveRehmannialeafmaterialsandtheExperimentEducationCenterforPharmaceuticalSciencesofZhejiangUniversityforhighperformanceliquidchromatographyHPLCanalysis.WegratefullyacknowledgethenancialsupportprovidedbytheNationalKeyTechnologyRDPrograminthe11thFiveYearPlanofChinaGrantNo.2006BAI06A08andtheImportantScienceTechnologySpecicProjectofZhejiangProvinceGrantNo.2008C03005.References1.L.S.LvJ.TangandC.T.HoSelectionandoptimizationofmacroporousresinforseparationofstilbeneglycosidefromPolygonummultiorumThunb.J.Chem.Technol.Biotechnol.8314222008.doi10.1002jctb.19642.Y.ZhaoB.ChenandS.Z.YaoSeparationof20S-protopanaxdioltypeginsenosidesand20S-protopanaxtrioltypeginsenosideswiththehelpofmacroporousresinadsorptionandmicrowaveassisteddesorptionSep.Purif.Technol.525332007.doi10.1016j.seppur.2006.06.0083.A.SchieberP.HiltandH.U.EndressAnewprocessforthecombinedrecoveryofpectinandphenoliccom-poundsfromapplepomaceInnov.FoodSci.Emerg.Technol4992007.4.Y.J.FuY.G.ZuW.LiuC.L.HouL.Y.ChenS.M.LiX.G.ShiandM.H.TongPreparativeseparationofvitexinandisovitexinfrompigeonpeaextractswithmacropo-rousresinsJ.Chromatogr.A113922062007.doi10.1016j.chroma.2006.11.0155.Y.J.FuY.G.ZuS.M.LiR.SunT.EfferthW.LiuS.G.JiangH.LuoandY.WangSeparationof7-xylosyl-10-deacetylpaclitaxeland10-deacetylbaccatinIIIfromtheremainderextractsfreeofpaclitaxelusingmacroporousresinsJ.Chromatogr.A11771772008.doi10.1016j.chroma.2007.11.0206.B.L.BianJ.YangW.HeZ.J.ZhangG.M.QingY..LiandD.M.Hu.ChinapatentCN1947757A2006.7.C.H.ChenT.Y.SongY.C.LiangandM.L.HuActeosideand6-O-AcetylacteosidedownregulatecelladhesionmoleculesinducedbyIL-1bthroughinhibi-tionofERKandJNKinhumanvascularendothelialcellsJ.Agric.FoodChem.571988522009.doi10.1021jf90283338.L.M.ReidC.P.ODonnellandG.DowneyRecenttechnologicaladvancesforthedeterminationoffoodauthenticityTrendsFoodSci.Technol.173442006.doi10.1016j.tifs.2006.01.0069.I.V.KovalenkoG.R.RippkeandC.R.HurburghDeterminationofaminoacidcompositionofsoybeansglycinemaxbynear-infraredspectroscopyJ.Agric.FoodChem.541034852006.doi10.1021jf052570u10.T.SunY.B.YingK.W.LiuandH.R.XuComparisonofchemometricsmethodsforassessinginternalqualityofpearson-lineusingvisiblenearinfraredtransmissionFigure6.Theadsorptionaandelutionbcurvesforacteo-sideNIRpredictedvaluesHPLCmeasuredvalues. 52RapidandQuantitativeDetectionofActeosideDuringPuricationofAdhesiveRehmanniaLeafExtracttechniqueinFoodProcessingAutomationConerenceProceedings2829June2008ProvidenceRhodeIslandUSA.ASABEStJosephMIchiganUSAPublicationNo.701P0508cd2008.11.R.M.BalabinandR.Z.SaevaNear-infraredNIRspectroscopyforbiodieselanalysisfractionalcomposi-tioniodinevalueandcoldlterpluggingpointfromonevibrationalspectrumEnerg.Fuels2523732011.doi10.1021ef200356h12.R.MBalabinandE.ILomakinaSupportvectormachineregressionSVRLS-SVManalternativetoneuralnetworksANNforanalyticalchemistryComparisonofnonlinearmethodsonnearinfraredNIRspectros-copydataAnalyst136817032011.doi10.1039c0an00387e13.D.PengG.H.ZhangK.N.DongandXiaLiLeastsquares-supportvectormachine-basedanalysisofnear-infraredspectrawithtechniquesofdimensionreductionandparameteroptimizationinProceedingsoftheSymposiumonPhotonicsandOptoelectronicSOPO.IEEEp.12010.doi10.1109SOPO.2010.550408214.X.C.GuoY.C.LiangandC.G.WuContentdetermina-tionbyPSO-basedLS-SVMregressioninProceedingofEighthInternationalConferenceonMachineLearningandCyberneticsp.10431215July2009BaodingChina2009.15.H.YangP.R.GrifthsandJ.D.TateComparisonofpartialleastsquaresregressionandmulti-layerneuralnetworksforquanticationofnonlinearsystemsandapplicationtogasphaseFouriertransforminfraredspectraAnal.Chim.Acta4891252003.doi10.1016S0003-26700300726-816.W.NieJ.G.LuoX.B.WangX.WanandL.Y.KongAninsightintoenrichmentandseparationofoleanane-typetriterpenoidsaponinsbyvariouschromatographicmaterialsSep.Purif.Technol.652432009.doi10.1016j.seppur.2008.10.04217.Y.T.WuL.C.LinJ.S.SungandT.H.TsaiDeterminationofacteosideinCistanchedeserticolaandBoschniakiarossicaanditspharmacokineticsinfreelymovingratsusingLCMSMSJ.Chromatogr.B844892006.doi10.1016j.jchromb.2006.07.01118.W.Z.LuandN.ModernInfraredSpectroscopyAnalyticalTechnology2ndEdn.ChinaPetrochemicalPressBeijingChina2006.19.M.BlancoJ.CoelloH.IturriagaS.MaspochandJ.PagesNIRcalibrationinnon-linearsystemsdiffer-entPLSapproachesandarticialneuralnetworksChemometr.Intell.Lab.Syst.50752000.doi10.1016S0169-74399900048-920.D.XiangJ.BerryS.BuntzP.GargiuloJ.CheneyY.JoshiB.WabuyeleH.Q.WuM.HamedA.S.HussainandM.A.KhanRobustcalibrationdesigninthepharma-ceuticalquantitativemeasurementswithnear-infraredNIRspectroscopyavoidingthechemometricpitfallsJ.Pharm.Sci.98311552009.doi10.1002jps.2148221.D.CozzolinoM.J.KwiatkowskiM.ParkerW.U.CynkarR.G.DambergsM.GishenandM.J.HerderichPredictionofphenoliccompoundsinredwinefer-mentationsbyvisibleandnearinfraredspectros-copyAnal.Chim.Acta513732004.doi10.1016j.aca.2003.08.06622.M.OttoandW.WegscheiderSpectrophotometricmulticomponentanalysisappliedtotracemetaldeter-minationsAnal.Chem.57631985.doi10.1021ac00279a02023.B.ParkJ.A.AbbottK.J.LeeC.H.ChoiandK.H.ChoiNear-infrareddiffusereectanceforquantitativeandqualitativemeasurementofsolublesolidsandrm-nessofdeliciousandgalaapplesTrans.ASAE4617212003.24.U.ThissenM.PepersB.UstunW.J.MelssenandL.M.C.BuydensComparingsupportvectormachinestoPLSforspectralregressionapplicationsChemometr.Intell.Lab.Syst.731692004.doi10.1016j.chemo-lab.2004.01.00225.M.J.AdamsChemometricsinAnalyticalSpectroscopy.RoyalSocietyofChemistryCambridgeUK1995.26.I.V.KovalenkoG.R.RippkeandC.R.HurburghMeasurementofsoybeanfattyacidsbynear-infraredspectroscopylinearandnonlinearcalibrationmethodsJ.Am.OilChem.Soc.834212006.doi10.1007s11746-006-1221-z27.T.NsT.IsakssonT.FearnandT.DaviesAUser-FriendlyGuidetoMultivariateCalibrationandClassication.NIRPublicationsPublicationsChichesterUK2002.28.L.GaoandS.H.RenPredictionofnitrophenol-typecom-poundsusingchemometricsandspectrophotometryAnal.Biochem.40521842010.doi10.1016j.ab.2010.06.03229.D.M.HimmelblauApplicationsofarticialneuralnet-worksinchemicalengineeringKoreanJ.Chem.Eng.173732000.doi10.1007BF0270684830.P.WilliamsandK.NorrisNear-InfraredTechnologyintheAgriculturalandFoodIndustries2ndEdn..AmericanAssociationofCerealChemistsStPaulMinnesotaUSA2001.31.J.A.KSuykensandJ.VandewalleLeastsquaressupportvectormachineclassiersNeuralProcess.Lett.92931999.doi10.1023A101862860974232.C.CortesandV.N.VapnikSupport-vectornetworksinMach.Learn.202731995.33.C.J.C.BurgesAtutorialonsupportvectormachinesforpatternrecognitionDataMin.Knowl.Discov.21211998.doi10.1023A100971592355534.T.V.GestelJ.A.K.SuykensB.BaesensS.ViaeneJ.VanthienenG.DedeneB.D.MoorandJ.VandewalleBenchmarkingleastsquaressupportvectormachineclassiersinMach.Learn.5452004.35.N.AyatM.CherietandC.SuenAutomaticmodelselectionfortheoptimizationofSVMkernelsPatternRecognit.3817332005.doi10.1016j.pat-cog.2005.03.011 Y.Jinetal.J.NearInfraredSpectrosc.21435320135336.F.FriedrichsandC.IgelEvolutionarytuningofmultipleSVMparametersNeurocomputing641072005.doi10.1016j.neucom.2004.11.02237.C.ChatelainS.AdamY.LecourtierL.HeutteandT.PaquetMulti-objectiveoptimizationforSVMmodelselectioninProceedingsNinthInternationalConferenceonDocumentAnalysisandRecognition.IEEEComputerSocietyWashingtonDCUSAp.4272007.38.M.ClercandJ.KennedyTheparticleswarmexplosionstabilityandconvergenceinamultidimensionalcom-plexspaceIEEETrans.Evol.Comput.6582002.39.Y.H.ShiandR.EberhartAmodiedparticleswarmoptimizerinProceedingsoftheIEEEWorldCongressonComputationalIntelligenceVol.104May1998AnchorageArkensasUSAp.691998.40.X.C.GuoJ.H.YangC.G.WuC.Y.WangandY.C.LiangAnovelLS-SVMshyper-parameterselectionbasedonparticleswarmoptimizationNeurocomputing7132112008.doi10.1016j.neucom.2008.04.02741.J.KennedyandR.EberhartParticleswarmoptimiza-tioninProceedingsoftheIEEEInternationalConferenceonNeuralNetworksVol.4.IEEEPressPiscatawayNJp.19421995.42.H.W.GeY.C.LiangY.ZhouandX.C.GuoAparticleswarmoptimization-basedalgorithmforjob-shopschedulingproblemsInt.J.Comput.Methods34192005.43.S.W.LinK.C.YingS.C.ChenandZ.J.LeeParticleswarmoptimizationforparameterdeterminationandfeatureselectionofsupportvectormachinesExpertSyst.Appl.3518172008.doi10.1016j.eswa.2007.08.08844.J.A.K.SuykensT.V.GestelJ.DeBrabanterB.DeMoorandJ.VandewalleLeastSquaresSupportVectorMachines.WorldScienticPublishingSingapore2002.45.M.StreeandP.F.G.VanHerpenB.VandeWaterbeemdL.A.VanderPolE.C.BeuveryJ.TramperD.E.MartensandM.ToftApracticalapproachforexplorationandmodelingofthedesignspaceofabacterialvaccinecul-tivationprocessBiotechnol.Bioeng.10434922009.doi10.1002bit.2242546.K.ZLiuM.ShiA.ManT.C.DembinskiandR.A.ShawQuantitativedeterminationofserumLDLcholesterolbynear-infraredspectroscopyVib.Spectrosc.382032005.doi10.1016j.vibspec.2005.04.00547.X.Y.NiuF.ShenY.F.YuZ.YanK.XuH.Y.YuandY.B.YingAnalysisofsugarsinChinesericewinebyFouriertransformnear-infraredspectroscopywithpartialleast-squareregressionJ.Agric.FoodChem.561672712008.doi10.1021jf800993e48.N.GierlingerM.SchwanningerandR.WimmerCharacteristicsandclassicationofFourier-transformnearinfraredspectraoftheheartwoodofdifferentlarchspeciesLarixsp.J.NearInfraredSpectrosc.121132004.doi10.1255jnirs.41549.J.J.KellyandJ.B.CallisNondestructiveAnalyticalpro-cedureforsimultaneousestimationofthemajorclassesofhydrocarbonconstituentsofnishedgasolinesAnal.Chem.6214441990.doi10.1021ac00213a01950.F.A.InR.LlarioS.GarriguesandM.delaGuardiaDevelopmentofaPLSbasedmethodfordeterminationofthequalityofbeersbyuseofNIRspectralrangesandsample-introductionconsiderationsAnal.Bioanal.Chem.382715492005.doi10.1007s00216-005-3343-951.J.W.HallandA.PollardNear-infraredspectrophotom-etryanewdimensioninclinicalchemistryClin.Chem.38916231992.52.A.SavitzkyandM.J.E.GolaySmoothinganddifferen-tiationofdatabysimpliedleastsquaresproceduresAnal.Chem.3616272003.doi10.1021ac60214a047 JOURNALOFNEARINFRAREDSPECTROSCOPY55ISSN0967-0335IMPublicationsLLP2013doi10.1255jnirs.1036AllrightsreservedThefrequentoccurrenceofpolymorphisminsolid-statepharmaceuticalsisanimportantissueindrugdevelop-mentandmanufacturing.1ChangesinpolymorphicformandpuritycansignicantlyinuencethephysicalandchemicalpropertiesofanactivepharmaceuticalingredientAPIwhichinturncanaffectbioavailabilityprocessabilityandshelf-life.Processing-inducedcrystalformconversionsduringmanu-facturingareaparticularconcern.Consequentlythemoni-toringofthepolymorphpurityofanAPIisnowpartofthequalityassuranceprocess.25ThesulfurdrugsulfathiazoleFigure1hasbeenextensivelyinvestigatedasamodelsysteminpolymorphismresearch.614Itisknowntoexistinvenon-solvatedpolymorphiccrystallineformsIV.Solventeffectsonthecrystallisationprocessleadingtodifferentpolymorphsorpolymorphicmixturesandsolid-statetrans-formationsofsulfathiazolepolymorphshavebeenthesubjectofseveralstudies.61114Thesestudiesrelyontheavailabilityoffastandaccuratemethodsforreal-timequantitativeAcomprehensivenearinfraredspectroscopicstudyofthelimitsofquantitativeanalysisofsulfathiazolepolymorphismPlMacFhionnghaileYunHuPatrickMcArdleandAndreaErxlebenSchoolofChemistryNationalUniversityofIrelandGalwayIreland.E-mailp.mcardlenuigalway.ieandrea.erxlebennuigalway.ieAcomprehensivestudyonthequanticationofsulfathiazolepolymorphsbynearinfraredNIRspectroscopycombinedwithmulti-variateanalysisisreported.Calibrationmodelsweredevelopedforeachofthesulfathiazolepolymorphsthatcanbeobtainedaspoly-morphicallypuresamplessulfathiazoleformsIIIIandVinbinarymixturesinthe015concentrationrange.PartialleastsquaresregressionPLSwasemployedanddifferentspectralpre-treatmentalgorithmsincludingmultiplicativescattercorrectionMSCstandardnormalvariateSNVtransformation1stderivativecalculationsandSNVcombinedwith1stderivativewereappliedtothedata.TheNIRmethodwascomparedwithX-raypowderdiffractionXRPDandwasfoundtogivemoreaccurateresultsinallcasesexceptonewhereXPRDwasmoreaccurate.FortheNIRmethodlimitsofdetectionwere0.5exceptforformVinIVmixtureswherethelimitofdetectionwas1.5.Limitsofquanticationrangedfrom1.1to4.6.WhenquantitiesofformsIIandIVwereaddedtosampleswhichwereanalysedusingthemodeldevelopedforformIIIinItheirpresencewasdetectedbyverylargepredictionerrors.AcomparisonofmultivariateandunivariateanalysisoftheNIRdatademonstratedthatingeneralmultivariatemethodsweresuperior.KeywordspolymorphismsulfathiazolepharmaceuticalscalibrationmodelspartialleastsquaresregressionIntroductionP.MacFhionnghaileetal.J.NearInfraredSpectrosc.2155662013Received28June2012Revised14December2012Accepted18December2012Publication23January2013Figure1.Chemicalstructureofsulfathiazole. 56LimitsofQuantitativeAnalysisofSulfathiazolePolymorphismsolid-stateanalysisofthesulfathiazolesystemthatallowthedetectionofsmallquantitiesofanotherpolymorph.WhenthestructuresofformsIIandIIIwereinitiallyreportedbyKrugerandGafner1516themeltingbehaviouroftheseformssuggestedatransitiontoformI.MorerecentlyZeitlerandco-workerscharacterisedthetransformationsofallfivesulfathiazolepolymorphsbyterahertzpulsedspectroscopyandDSCandreportedsolid-statetransformationsofformsIIIIIandIVtoformIatvaryingtemperatures.11ThecrystallatticesofformsIIIIIandIVareverysimilar.AnanalysisusingtheprogramXPachasshownthatformsIIandIVcontainthesamemonolayerswiththedifferencebetweenthembeingthatthemonolayersareslippedinformIVrelativetoformIIandthatformIIIisacombinationofthebilayersofIIandIV.17TheepitaxialgrowthofformIVonformIIcrystalshasalsobeenobservedinsolutionandthenucleationprocesshasbeenassociatedwiththeirstructuralsimilarities.18ItisthusnotsurprisingthatitisdifculttocrystallisepuresamplesofformsIIIIIandIVfromsolution.TraditionallyandnowforalmostacenturyX-raypowderdiffractionXRPDhasbeenthegoldstandardforquantitativephaseanalysisandthequanticationofpolymorphicmixtures.HoweveraprecisequantitationusingXRPDrequirestheaccu-ratemeasurementoftheintensitiesandareasofdiffractionpeakswhicharestronglyinuencedbysampleinstrumentandsamplepreparationparameters.19Althoughsampleprop-ertiesandpreparationaspotentialsourcesoferrorcanbepartlyovercomebywholepowderpatternttingmethods2022vibrationalspectroscopictechniquessuchasRamannearinfraredNIRandattenuatedtotalreflectanceinfraredspectroscopyareemergingmoreandmoreastechniquesofchoiceforquantitativesolid-stateanalysis.23InparticularRamanandNIRspectroscopyofferadvantagesforreal-timestudiesofbatchcrystallisationsbecausetheycanbereadilycoupledtobre-opticprobesarefastandnon-destructiveandrequireminimalsamplepreparation.HoweverthelargersizeoftheirradiatedsamplevolumeusedbyNIRinstrumentsdoesgivetheNIRmethodanadvantageoverRamanspectroscopy.Traditionaluni-orbivariatemethodsbasedontheanalysisoftheabsorbanceatsinglecharacteristicwavelengthsintheNIRspectrumhavebeensuccessfullyappliedforthequanticationofbinarypolymorphicmixturesofsolid-statepharmaceuticals.24Howevermoresophisticatedmultivariatemethodssuchaspartialleast-squaresPLSregressionthatutilisespectralinformationfromthewholespectrumorselectedspectralregionsaremoreeffectiveespeciallywhenthespectraldifferencesbetweenpolymorphsaresmall.Nowadaysmultivariatetechniquesoftenresultinmorereliableandrobustcalibrationmodelsandaregenerallyconsideredsuperiortouni-orbivariateapproaches.2528Literaturereportsonthequanticationofsulfathiazolepoly-morphshavefocusedtodateonthequanticationofformsIandIIILuneretal.useduni-andbivariateapproachesforthequanticationofbinarymixturesofformsIandIIIbyNIRspectroscopyandXRPDandobtainedaccuratelinearregres-sionmodelswithhighcorrelationvaluesR2andlowstandarderrorsofcalibration.29Patelandco-workersevaluatedtheabilityofNIRspectroscopycombinedwithunivariatemultiplelinearregressionandPLSmethodstoaccuratelydeterminelowcontent05offormIinformIII.24Pllnenetal.haveusedDiffuseReectanceInfraredFourierTransformInfraredDRIFT-IRspectroscopyanddevelopedPLSmodelstoquantifyformIandIIIinpolymorphicmixturesofformIIIIIIandV.30Inthispaperwehaveattemptedtherstcomprehensivelowcontentsulfathiazolepolymorphanalysis.Theaimofthepresentstudywastoassessthelimitsofpolymorphanalysisforthesulfathiazolesystemandtofurtherexploretheaccu-racyreliabilityandrobustnessofNIRspectroscopicmethodsforthedetectionoflowlevelsofacrystallinephase.MaterialsandmethodsMaterialsSulfathiazoleformIIIwaspurchasedfromSigmaAldrichwithapurityof98andwasusedasreceived.FormsIandVwerepreparedaccordingtoliteratureprocedures.8929TheidentityandpurityofthepolymorphswasconfirmedbyXRPDanddifferentialscanningcalorimetryDSC.Thecrystalstruc-turesofthepolymorphswereverifiedbycomparingtheexperimentalXRPDpatternswiththetheoreticaldiffracto-gramscalculatedfromatomiccoordinatesobtainedfromtheCambridgeCrystallographicDataCentreCCDCusingOscailSoftwareSupplementaryInformation.31Structureswithreferencecodessuthaz01suthaz02andsuthaz05wereusedtogeneratethetheoreticaldiffractionpatternsofformsIIIIandV.ThebulksamplesofthepolymorphsweremonitoredbyNIRspectroscopyandXRPDtoconrmthateachpolymorphwasphysicallystableoverthecourseofthestudy.Tomini-miseparticlesizeeffectsthebulksamplesofeachformweremilledunderthesameconditionsusingaplanetarymicromillPulverisette7FritschGmbHIdar-ObersteinGermany.XRPDandNIRmeasurementsdemonstratedthatallformswerestabletowardsgrindingandneitheramorphisationnorpolymorphchangeswereobserved.PreparationofpolymorphicmixturesSixsetsofbinarypolymorphicmixturesofformsIIIIIVandIIIVcontaining0123456789101113and15oftherespectiveminorcomponentwereprepared600mgtotalforeachmixture.Priortoanalysisthephysicalmixtureswereagitatedrepeatedlyusingavortexmixer.AllmixturesweresealedwithTeon-linedcapsandkeptinadesiccatoratambienttemperature.Inordertoavoidpossiblesystematicchangescausedbyinstrumentalandorenvironmentaluctuationssampleswereanalysedinarandomorderimmediatelyafterpreparation.DifferentialscanningcalorimetryDSCmeasurementstoconrmtheidentityandpurityofthesulfathiazolepolymorphswerecarriedoutonaSTA625thermalanalyserfromRheometricScienticPiscatawayNJUSAin P.MacFhionnghaileetal.J.NearInfraredSpectrosc.215566201357openaluminiumcruciblesataheatingrateof10Cmin1.Thetemperaturerangewasbetween25Cand250C.Nitrogenwasusedasthepurgegas.AnindiumstandardwasusedforDSCcalibration.X-raypowderdiffractometryX-raypowderdiffractiondatawerecollectedonaSiemensD500powderdiffractometerMunichGermanywhichwasfittedwithadiffractedbeammonochromator.Diffractionpatternswererecordedbetween5and402qusingCuKaradiationl1.54withstepsof0.05anda2scountingtimeperstep.NearinfraredspectroscopyNIRspectrawererecordedinglassvials15mm45mmonaPerkinElmerSpectrumOnespectrometerWalthamMAUSAttedwithanNIRreectanceattachment.Spectrawerecollectedoverthe10000cm14000cm1rangewithareso-lutionof8cm1using32co-addedscans.Allsampleswerere-sampledthreetimesandthemeanspectrumwasused.DataanalysisMultivariatedataanalysiswascarriedoutusingthemulti-variatedataanalysissoftwareTheUnscramblerv.9.8CamoNorway.Pre-treatmentmethodsincludedmultiplicativescattercorrectionMSCstandardnormalvariateSNVtransformationrstderivativecalculationsandrstderiva-tivecalculationscombinedwithSNVfortheNIRdataandSavitzkyGolaysmoothingandrstderivatisationfortheXRPDdata.Forunivariateanalysisalinearbaselinecorrectionandsecondderivativecalculationswereappliedaspre-treatmentmethods.Savitzky-Golay1stand2ndderivativecalculationsandsmoothingwereperformedwithawindowsizeof7XRPDpatternsand11pointsNIRspectraanda2ndorderpoly-nomial.ThegapandsegmentsizesusedforsecondderivativecalculationswereNorrisGapgapsize7andGap-segmentgapsize7segmentsize3.Theloadingplotswereusedtoidentifysuitablespectraland2qregionsforbuildingthecali-brationmodels.Differentwavenumberand2qrangeswereevaluatedandthebest-performingmodelwasselected.Thespectroscopicdataweremean-centredbeforeapplyingpartialleast-squaresPLS.Everysetofmeasurementswassplitintoacalibrationset01235781011and15andapredictionset469and13.TheoptimalnumberofPLSfactorswasdeterminedusingaleave-one-outcross-validationprocedureorbyndingtheminimumrootmeansquareerrorofpredictionRMSEP.32Forcomparingthedifferentpre-processingmethodstheroot-mean-squareerrorsofcalibrationRMSECcross-validationRMSECVandRMSEPwereused.Thesearedenedas21niiiyyRMSEn-whereyiisthereferencevalueyithecalculatedvalueandnisthenumberofsamples.ResultsanddiscussionConstructionofanalyticalmodelsInanattempttoincludeallvepolymorphsinouranalysisthesyntheticmethodsforobtainingthepureformswereanalysed.ItwasnotpossibletoobtainpolymorphicallypuresamplesofformsIIandIV.ForexampleallattemptstocrystalliseformIIfrommethanolornitromethanefollowingliteraturereports1333onthesolvent-controlledcrystallisa-tionofsulfathiazolepolymorphsyieldedmixtures.SamplesobtainedfromethanolconsistedmainlyofformIIwhilecrystallisationfromwatergavesamplescontainingformIVasthemaincomponent.TheobservationthatformsIIIIIandIVusuallycrystallisetogetherhasalsobeenmadebyotherauthors.10Thereasonforthisislikelytheclosesimilarityoftheircrystallattices.ThusaccuratebinarymodelmixturescouldonlybepreparedforformsIIIIandV.Thelimitsthatthisplacesonthecomprehensivenessoftheanalysisarediscussedbelow.NIRspectroscopyDuetodifferencesinhydrogenbondingformsIIIIandVshowdistinctvibrationalspectrawhichallowforscreeningandquantitationbyIRandNIRspectroscopy.243436TheNIRspectraofpolymorphsIIIIandVdiffermostinthe7050cm16200cm1rangewhichencompassestheregionoftherstovertonesfortheNHstretchingvibrationsandintheregionfrom5150cm1to4800cm1wherethecombinationbandsofNHstretchingandNH2bendingvibrationsappearFigure2.CalibrationmodelsweredevelopedforeachforminmixturesofformsIIIIIVandIIIVinthe015concen-trationrangeusingaPLSregression.WhenthefollowingspectralrangeswereselectedPLSmodelswithlowrootmeansquareerrorRMSEvalueswereobtained69706540cm1and51205000cm1for015formIinformIII65046230cm161326060cm1and49644850cm1for015formIIIinformI69606830cm150845050cm1and45584516cm1for015formIinformV67506300cm1and61606046cm1for015formVinformI69006520cm1and51004990cm1for015formVinformIIIand64856174cm1and49824862cm1for015formIIIinformV.ForformIinIIIformIIIinIformIinVandformVinIIItheselectedrangeencompassesbandsspecictotherespectiveminorcomponent.LowlevelsofformIinIIIImixturesarevisibleasweakbandsat6870cm16730cm15078cm15058cm1and5036cm1whilelowlevelsofformIIIcanbeidentiedbycharacteristicnewbandsataround6400cm16126cm1and6110cm1andclearchangesaround4900cm1.SimilarlylowlevelsofformVinIIIVmixturesgiverisetoweaknewbandsat6800cm15076cm15048cm1and5010cm1.LowlevelsofformIinIVmixturescanbevisuallyidentifiedbyacharacteristicformIbandwhichappearsasashoulderataround6870cm1.LowlevelsofformIIIinVresultinclearchangesaround6400cm1and4900cm1.Incontrastitisclearfromthevisualinspection 58LimitsofQuantitativeAnalysisofSulfathiazolePolymorphismoftheNIRspectraofIVmixturescontaining015formVthattherearenonewspectralfeaturesassociatedwithincreasingamountsofformVFigure2.Indeedthecombi-nationbandregionsofIandVarerathersimilar.HoweverincreasingamountsofformVgiverisetosmallincreasesintheintensityofthebandat6140cm1.TheRMSECRMSECVandRMSEPvaluesandthenumberofPLSfactorsforthecalibrationmodelsdevelopedfortheNIRmeasurementsarereportedinTable1.Pre-treatmentoftherawdataenhancedthemodelperformancesandreducedthenumberofPLSfactorsrequired.UsingSNVorMSCaspre-processingtechniquesgavethebestresultsforthe70006800660064006200600058005600AbsorbanceWavenumbercm-1acb500048004600440042004000AbsorbanceWavenumbercm-170006800660064006200600058005600AbsorbanceWavenumbercm-1500048004600440042004000AbsorbanceWavenumbercm-170006800660064006200600058005600AbsorbanceWavenumbercm-1500048004600440042004000AbsorbanceWavenumbercm-1Figure2.70005600cm1and51504000cm1rangeoftheNIRspectraofbinarymixturesofsulfathiazolepolymorphs.aFormsIIII015formIsolidlines015formIIIdashedlines.bFormsIV015formIsolidlines015formVdashedlines.cFormsIIIV015formIIIsolidlines015formVdashedlines.Black0red2green5blue8cyan10magenta15.Thisgureisincolourintheon-lineversion. P.MacFhionnghaileetal.J.NearInfraredSpectrosc.215566201359quanticationoflowlevelsofonepolymorphinIIIIandIVbinarymixtures.InallcasesMSCandSNVgaveverysimilarresults.ThisisoftenobservedastheequationsforMSCandSNVhavethesameformandarewidelyregardedasexchangeable.3738MSCandSNVremoveoffsetsandslopesinthespectracausedbythelightscatteringintrinsictosolidsamples.TheadvantageofSNVisthatitisappliedtoanindividualspectrumwhereasMSCusesareferencespectrum.Thebestmodelsforthedeterminationof015formVinformIIIandof015formIIIinformVwereobtainedwhenrstderivativecalculationswereappliedorwhenSNVandrstderivatisationwerecombined.Overallthebest-performingcalibrationmodelresultedforthelowcontentdetermi-nationofformIIIinIIIImixtureswithRMSECRMSECVandRMSEPvaluesof0.2010.340and0.298andtwoPLSfactors.PlotsofpredictedvsactualcontentforthesixdifferentbinarymixtureswhichallshowexcellentlinearcorrelationsaredisplayedinFigure3.TheNIRspectraofformsIIandIVdifferfromthespectrumofformIinthatIIandIVgiverisetoabroadbandat6400cm1thatisnotobservedforI.HoweveraspuresamplesofformsIIandIVarenotavailablecalibrationmodelsforIIIandIIVmixturescannotbeobtained.ThepossiblepresenceofformsIIandIVinrealsamplesandtheeffectthiswouldhaveonpredictionsmadeusingmodelsconstructedfromformsIIIIandVwasexaminedvideinfra.DifcultiesaretobeexpectedsinceformsIIandIVcloselyresembleformIIIinallspectralandphysicochemicalbehav-iour.10IndeedtheNIRspectraofformsIIIandIVexhibitonlyaminordifferenceintheintensityofthebandat6112cm1Figure4whiletheNIRspectraofformsIIandIIIshowsomedifferencesinthe67506000cm1and51004700cm1regionsFigure4.TheclosesimilarityofthevibrationalspectraofformsIIIIIandIVisduetothefactthattheH-bondingpatterninformIIIineffectcombinestheH-bondingmotifsofformsIIandIV.6ThustheNIRspectrumofformIIIisvirtuallythesumofthoseofformsIIandIII.11X-raypowderdiffractionChangesintheXRPDpatternsduetolowlevelsofonepoly-morphareshowninFigure5forIIIIIVandIIIVmixtures.PerformancecharacteristicsofthemodelsgeneratedusingmultivariateXRPDmethodsaresummarisedinTable2.ThebestresultsRMSECRMSECVandRMSEPvaluesof0.2140.417and0.240havingjusttwoPLSfactorswereobtainedforthequantificationof015formIinformIIIwithPLSmodelsbuiltontherstderivativeofthe2qregions15.616.117.017.9and18.619.0.ThreepeaksspecictoformIat16.017.7and18.9lieintheseregions.Incontrastthecali-brationmodelforthedetectionoflowlevelsofformIIIinformIwhichgiverisetocharacteristicpeaksat15.4and25.6hadBinarymixturePre-processingmethodPLSfactorsRMSECRMSECVRMSEPPLSfactorsRMSECRMSECVRMSEP015formIa015formIIIbFormsIIIIRawdata30.1760.3370.39230.1200.2760.413SNV10.2990.3710.26620.2010.3400.298MSC10.2990.3720.26620.2030.3460.2961stder.10.1960.2530.46710.2920.4950.452SNV1stder.10.1810.2300.63710.2840.4480.313015formIc015formVdFormsIVRawdata40.3831.0390.72440.4211.0160.896SNV10.4600.6080.48520.5750.7670.446MSC10.4580.6050.48520.5750.7670.4491stder.30.3440.6600.58810.6440.8120.932SNV1stder.20.4880.7480.68710.6640.8080.994015formIIIe015formVfFormsIIIVRawdata30.3760.7410.31030.4460.7440.385SNV10.4140.5150.40430.4390.7260.481MSC10.4100.5090.40130.4440.7140.4761stder.20.3970.5530.27920.4480.6720.344SNV1stder.10.4020.5390.27920.4490.7040.377a69706540cm1and51205000cm1range.b6504623061326060and4964cm14850cm1range.c6960683050845050and45584516cm1range.d67506300cm1and61606046cm1range.e64856174cm1and49824862cm1range.f69006520cm1and51004990cm1range.Table1.PerformancecharacteristicsofmultivariateNIRmethods. 60LimitsofQuantitativeAnalysisofSulfathiazolePolymorphismsignicantlylargerRMSEvaluesTable2.LikewisetheresultsindicatedthatXRPDwasclearlyinferiortoNIRspectroscopyforthequanticationoflowcontentofformVinIVandIIIVmixtures.TheXRPDpatternofformVshowspeaksat20.3and23.3whicharespecictothispolymorph.8Signicantinten-sityvariationshavebeenobservedforthepeakat23.3duetopreferredorientationeffectsoneofthemainfactorsthataffecttheaccuracyofquantitativeanalysisbyXRPD.Thefactthattheregionofthecharacteristicpeakat23.3isnotanalyticallyusefulmayexplainthepoormodelperformancesforIVandIIIVmixturescontaininglowlevelsofformV.AgainincontrastsatisfactorymodelswereobtainedformixturescontaininghighlevelsofformV.InfacttheXRPDmethodgaveslightlybetterresultsthanNIRspectroscopyforthequanticationoflowformIcontentinthepresenceofformV.Useoftheregionschar-acteristicofformsIandIII17.217.921.822.5and14.715.721.822.6respectivelyledtomodelsthathadrelativelygoodpredictiveabilityforlowlevelsofthesepolymorphsintheirbinarymixtureswithformV.ComparisonofmultivariateandunivariatemethodsPateletal.havecomparedunivariatemultiplelinearregressionandPLSmethodsforthequanticationof05sulfathiazoleformIinformIIIbyNIRspectroscopy.24Usingthenormalisedsecondderivativevalueatasinglewave-lengththeywereabletoconstructanaccurateunivariate02468101214160246810121416CalibrationsetTestsetaPLSpredictedpercentageActualpercentagebyweight02468101214160246810121416CalibrationsetTestsetbPLSpredictedpercentageActualpercentagebyweight02468101214160246810121416CalibrationsetTestsetcPLSpredictedpercentageActualpercentagebyweight02468101214160246810121416CalibrationsetTestsetdPLSpredictedpercentageActualpercentagebyweight02468101214160246810121416CalibrationsetTestsetePLSpredictedpercentageActualpercentagebyweight02468101214160246810121416CalibrationsetTestsetfPLSpredictedpercentageActualpercentagebyweightFigure3.Predictedvsactualpercentageforbinarymixturesanalysedinthe015compositionrangebyNIRspectroscopy.a015formIinformIIISNVpre-processing69706540cm1and51205000cm1range.b015formIIIinformISNVpre-processing65046230cm161326060cm1and49644850cm1range.c015formIinformVSNVpre-processing69606830cm150845050cm1and45584516cm1range.d015formVinformISNVpre-processing67506300cm1and61606046cm1range.e015formIIIinformVSNVand1stderivative64856174cm1and49824862cm1range.f015formVinformIII1stderiva-tive69006520cm1and51004990cm1range.Thesolidlinesrepresentthedatatwithalinearregressionmodel.7000675065006250600050004750450042504000AbsorbanceWavenumbercm-1abcFigure4.NIRspectraofsulfathiazoleaformIIIandcrystalsobtainedfrombethanolandcwater.CrystalsobtainedfromethanolaremainlyformIIwhilesamplesobtainedfromwatercontainformIVasthemaincomponent. P.MacFhionnghaileetal.J.NearInfraredSpectrosc.215566201361calibrationmodelwithstandarderrorofcalibrationSECandstandarderrorofpredictionSEPvaluesof0.075and0.069.IncontrastmultivariatePLSmethodswhilegivingSECvaluescomparabletotheunivariatecalibra-tionmodelresultedinlesspredictiveless-robustmodels.Theseauthorshavesuggestedthatoverttingortheiruseofalownumberofcalibrationstandardsarepossiblereasonsforthepoorperformanceofmultivariatemodels.24AsdescribedaboveapplyingPLSregressiontoourNIRdatainthe015rangeresultedinsatisfactoryRMSECandRMSEPvaluesforallsixbinarymixturesstudiedincludingmixturescontaininglowlevelsofformIinformIII.InallcasesaclosecorrespondenceofRMSECandRMSEPwasfoundindicatinggoodpredictiveabilityandrobustcalibra-tionmodels.Tocomparemultivariateandunivariatemethodsingreaterdetaildifferentunivariatemodelswereconstructedusingdifferentpre-treatmentalgorithmsTable3.ForthelowleveldeterminationofformIinIIIunivariatemethodscombinedwithbaselinecorrectionSNVandsecond-derivativetransfor-mationgaveRMSECvaluesjustslightlybetterthanmultivariatemethods.ThebestresultswereobtainedwhentheNIRspectrawereSNVcorrectedRMSEC0.224RMSEP0.178.IncontrasttoPatelsanalysis24univariateanalysisofour2ndderivativetransformedNIRdataSupplementaryInformationgavehigherRMSEPvaluesthanthebest-performingmulti-variatemodel0.42forunivariatevs0.27formultivariate.Wealsohavetonotethatnoneofthepre-processingmethodsappliedtoourdataledtoRMSECandRMSEPvaluesaslowasthoseobtainedbyPateletal.Howeverdespitethepoorerperformanceofourunivariatemodelanddespitethefactthatourbestperformingmodelisbasedonadifferentpre-treatmenttechniqueandthusisnotdirectlycomparabletotheliteraturedataourresultsalsosuggestthatmultivariatemethodsdonotofferanyadvantageoverunivariatemethodsforthequanticationoflowlevelsofformIinformIII.Thesameistrueforthedeterminationof015formIIIinformI.IncontrastforthelowcontentdeterminationofformIinVformVinIformIIIinVandformVinIIImultivariatemethodswereclearlysuperiortounivariatemethodswithregardtotherobustnessandpredictiveabilityofthecalibrationmodels.RMSEPvaluesforthebest-performingunivariatemodelswere0.748formIinV1.057formVinI0.474formIIIinVand0.457formVinIIIcomparedto0.485formIinV0.446formVinI0.279formIIIinVand0.344formVinIII.LimitsofdetectionandquanticationTofurtherevaluatetheperformanceoftheanalyticalmethodsthelimitsofdetectionLODandlimitsofquanticationLOQhavebeencalculated.LODandLOQvaluesweredeterminedusingblanksamplesandLOD3.3sssstandarddevia-tionsslopeofthecalibrationplot39andLOQ10xss.19Thecalibrationcurveforthebestperformingmultivariatemodelforeachbinarymixturewasused.LODandLOQvaluesarereportedinTable4.LOQforformsIIIIandVinthesixbinarymixturesstudiedrangefrom1.1to4.6fortheNIRmethod.LODare0.5forformIinIIIformIIIinIformIinVformVinIIIandformIIIinV.AsignicantlyhigherLODwasobtainedforformVinbinarymixtureswithformI1.5.XRPDgaveLODvaluesbetween0.7and2.8andLOQvaluesbetween2.0and8.5.AsdescribedabovethecombinationbandregionsofformsIandVarerathersimilarandnonewbandsappearintheNIRspectraofIVmixtureswithincreasingamountsofformV.ThismayexplaintherelativelyhighLODforformVinformI.Incontrastinallothercasestheminorcomponentcan141618202224260500100015002000250030003500Intensity2thetaoa161820222426050010001500200025003000Intensity2thetao141618202224260500100015002000250030003500Intensity2thetaobcFigure5.SmoothedXRPDpatternsofbinarymixturesofsulfathiazolepolymorphs.aFormsIIII015formIsolidlines015formIIIdashedlinesbsulfathiazoleformsIV015formIsolidlines015formVdashedlinescsulfathiazoleformsIIIV015formIIIsolidlines015formVdashedlines.Black0red5green10blue15.Thisgureisincolourintheon-lineversion. 62LimitsofQuantitativeAnalysisofSulfathiazolePolymorphismbevisuallyidentiedbycharacteristicbandsandorclearchangesintheselectedspectralregionsvidesupra.TheresultsforthespectroscopicmethodcomparewellwiththosereportedforotherbinarypolymorphanalysesandconfirmthesuitabilityofNIRmethodsforthelowcontentquanticationofsulfathiazolepolymorphsinbinarymixtures.XRPDmethodsreportedintherecentliteraturetypicallygiveLODvaluesforpolymorphsinbinarymixturesofaround1.1932OlanzapineisoneexamplewhereaLODforitspolymorphIaslowas0.4hasbeenobtainedusingaBinarymixturePre-processingmethodRMSECRMSEPRMSECRMSEP015formI015formIIIFormsIIIISNV0.224a0.178a0.336b0.623bBaselinecorrection0.182c0.436c0.284b0.342bSavitzky-Golay2ndder.0.264d0.456d0.570b0.774bNorrisGap2ndder.0.260d0.419d0.482b0.417bGap-segment2ndder.0.252d0.415d0.383b0.308b015formI015formVFormsIVSNV1.019e0.748e1.666f1.915fBaselinecorrection0.911e1.535e3.061f2.869fSavitzky-Golay2ndder.1.442g2.684g0.791h1.057hNorrisGap2ndder.0.919g1.481g0.854h1.063hGap-segment2ndder.0.795g1.220g0.789h1.125h015formIII015formVFormsIIIVSNV0.434i0.474i0.716j0.777jBaselinecorrection0.882k1.206k0.6740.856Savitzky-Golay2ndder.0.571l1.085l0.608a0.459aNorrisGap2ndder.0.527l0.801l0.571a0.457aGap-segment2ndder.0.552l0.740l0.594a0.533aa5076cm1b6876cm1c5080cm1d6410cm1e6870cm1f6866cm1g6140cm1h4850cm1.i4920cm1.k6400cm1.6370cm1.m5046cm1.n6800cm1Table3.PerformancecharacteristicsofunivariateNIRmethods.BinarymixturePre-processingmethodPLSfactorsRMSECRMSECVRMSEPPLSfactorsRMSECRMSECVRMSEP015formIa015formIIIbFormsIIIIRawdata10.2340.3190.45020.4990.8170.602Smoothing10.2380.3580.46620.5570.7700.5511stderiv.20.2140.4170.24020.6470.9450.421015formIc015formVdFormsIVRawdata10.3870.5090.46210.8551.0640.791Smoothing10.3720.4890.36910.8541.0530.7711stderiv.20.4780.7761.08340.2901.3690.428015formIIIe015formVfFormsIIIVRawdata10.6600.9120.53720.3550.8421.435Smoothing10.6640.9330.64430.3550.9881.4841stderiv.10.7931.0750.72030.8251.6562.820a15.616.117.017.9and18.619.0range2qb14.715.7and25.225.8range2qc17.217.9and21.822.5range2q.d20.020.6range2qe14.715.7and21.822.6range2qf15.716.1and20.220.5range2q.Table2.PerformancecharacteristicsofmultivariateXRPDmethods. P.MacFhionnghaileetal.J.NearInfraredSpectrosc.215566201363univariateXRPDmethod.40RecentlyithasbeenshownthatchemometricIRmethodsaresuitabletoquantifylowlevelsofclopidogrelbisulfatepolymorphswiththeLODbeing0.5forformII.41Uni-andmultivariateRamanmethodshavebeenreportedtogiveLODvaluesforbinarypolymorphicmixturesofaround12.334142IthastobenotedthatPateletal.obtainedaLODof0.06forsulfathiazoleformIinformIIIapplyingNIRspectroscopyandunivariatemethodsasdescribedabove.24TestoftheformIIIinIIIImixturesmodelonsamplescontainingcontaminantformsIIandIVWehavestatedabovethatitisdifculttoobtainpuresamplesofformsIIandIVhoweverfollowingrecommendationsinarecentreview43relativelypuresamplesofIIandIVwereobtainedbycrystallisationofsulfathiazolefromsaturatedethanolandaqueoussolutionrespectively.TotestthelowcontentformIIIinIIIImodelinthepresenceofformsIIorIVascontaminantpolymorphssamplescontainingaddedformsIIandIVwerepreparedsamplesVtoVIIItogetherwithrefer-encesamplescontainingformsIandIIIonlysamplesItoIVFigure6andTable5.NIRspectraofthesesampleswereusedtopredicttheformIIIcontentusingthelowcontentformIIIinIIIImodel.ThepredictionresultsshowthattheadditionofformsIIandIVhavetwoeffects.FirstformsIIandIVpartlymimicformIIIandthepredictedformIIIcontentishigherthanthetrueformIIIcontentshowninredinFigure6.SecondthepresenceofformsIIandIVleadstopredictionswithmuchlargerpredictionerrors.Thusthesizeofthepredictionerrorscanbeusedtoidentifyinappropriateuseofthemodel.BinarymixtureNIRspectroscopyXRPDLODwtLOQwtLODwtLOQwtFormsIIIIFormIFormIII0.430.451.301.360.761.202.303.64FormsIVFormIFormV0.451.501.364.550.942.802.858.49FormsIIIVFormIIIFormV0.400.361.211.090.661.662.005.03Table4.LODandLOQformultivariateNIRspectroscopyandXRPDmethods.SampleActualcontentformIActualcontentformIIActualcontentformIIIActualcontentformIVPredictedcontentformIIIaI96.04.04.20.6II94.06.06.30.6III91.09.09.20.4IV87.013.012.60.4V90.010.06.02.6VI85.010.05.010.42.3VII90.010.09.01.5VIII85.05.010.013.21.4aLowcontentformIIIinIIIImodelIIIIIIIVVVIVIIVIII0246810121416SamplenameContentofformIIIFigure6.ContentofformIIIintestsamplesIVIIIcf.text.Redactualcontentblackpredictedcontentwitherrorbars.Thisgureisincolourontheon-lineversion.Table5.ActualandpredictedcontentofformIIIintestsamples. 64LimitsofQuantitativeAnalysisofSulfathiazolePolymorphismConclusionsInthepresentstudycalibrationmodelsforbinarymixturesofallknownsulfathiazolepolymorphsthatcanbeobtainedaspolymorphicallypurecrystallinephaseshavebeenconstructed.TothebestofourknowledgecalibrationmodelsforbinarymixturesofformsIVandformsIIIVorforlowlevelsofformIIIinformIhavenotbeenpreviouslydescribed.LowlevelsofsulfathiazoleformsIIIIandVinbinarymixturescanbeaccuratelyquantiedbyNIRspectros-copywithLODandLOQvaluesrangingfrom0.4to1.5and1.14.6respectively.Howevercarefulselectionofanalyti-callyusefuldataregionsisrequired.AlthoughsatisfactorymodelscouldbedevelopedforthelowcontentquanticationbyXRPDitisclearthatNIRspectroscopyexhibitshigheraccuracyonthebasisofRMSEvaluesexceptforthedetermi-nationoflowlevelsofformIinIVmixtureswheretheXRPDmethodgaveslightlylowerRMSEvalues.FurthermoreNIRspectroscopygivessignicantlylowerLODandLOQvaluesthanXRPD.CombinedwiththeconvenientinglassbottlesamplingmethodthepresentresultsprovidefurtherinsightintotheeffectivenessofNIRspectroscopyforpolymorphanalysis.WhenoneofthemodelswastestedusingsampleswhichcontainedformsIIandIVwhichwerenotincludedinthemodelsthepredictionresultsshowalargeincreaseinthepredictionerror.Univariatemethodswerecomparedwithmultivariatemethodsandwhiletheformerallowfortheaccurateandreliableanalysisoftwoofthesixbinarymixturesstudiedmultivariatemethodswerefoundtobethemoregenerallyapplicableapproach.AcknowledgementThisworkwasfundedbyScienceFoundationIrelandaspartoftheSolidStatePharmaceuticalClusterGrantNo.07SRCB1158.MrDermotMcGrathisthankedforDSCmeasurements.References1.H.G.BrittainPolymorphismandsolvatomorphism2007J.Pharm.Sci.98516172009andrefs.therein.doi10.1002jps.215182.G.A.StephensonR.A.ForbesandS.M.Reutzel-EdensCharacterizationofthesolidstatequantitativeissuesAdv.DrugDeliv.Rev.481672001.doi10.1016S0169-409X0100099-03.InternationalConferenceonHarmonizationICHSpecicationsTestProceduresandAcceptanceCriteriaNew.DrugSubst.New.DrugProductsChem.Subst.October61999.4.S.R.VippaguntaH.G.BrittainandD.J.W.GrantCrystallinesolidsAdv.DrugDeliv.Rev.48132001.doi10.1016S0169-409X0100097-75.S.ByrnR.PfeifferM.GaneyC.HoibergandG.PoochikianPharmaceuticalsolidsastrategicapproachtoregulatoryconsiderationsPharm.Res.1279451995.doi10.1023A10162419274296.N.BlagdenR.J.DaveyH.F.LiebermanL.WilliamsR.PayneR.RobertsR.RoweandR.DochertyCrystalchemistryandsolventeffectsinpolymorphicsystemssulfathiazoleJ.Chem.Soc.FaradayTrans.9410351998.doi10.1039a706669d7.M.LagasandC.F.LerkThepolymorphismofsulfathiazoleInt.J.Pharm.8111981.doi10.10160378-51738190023-58.J.AnwarS.E.TarlingandP.BarnesPolymorphismofsulfathiazoleJ.Pharm.Sci.7843371989.doi10.1002jps.26007804169.F.C.ChanJ.AnwarR.CernikP.BarnesandR.M.WilsonAbinitiostructuredeterminationofsulfathiazolepolymorphVfromsynchrotronX-raypowderdiffractiondataJ.Appl.Crystallogr.324361999.doi10.1107S002188989801723310.D.C.ApperleyR.A.FlettonR.K.HarrisR.W.LancasterS.TavenerandT.L.ThrelfallSulfathiazolepolymorphismstudiedbymagic-anglespinningNMRJ.Pharm.Sci.881212751999.doi10.1021js990175a11.J.A.ZeitlerD.A.NewnhamP.F.TadayT.L.ThrelfallR.W.LancasterR.W.BergC.J.StrachanM.PepperK.C.GordonandT.RadesCharacterizationoftemperature-inducedphasetransitionsinvepoly-morphicformsofsulfathiazolebyterahertzpulsedspectroscopyanddifferentialscanningcalorimetryJ.Pharm.Sci.951124862006.doi10.1002jps.2071912.J.E.AndersonS.MooreF.TarczynskiandD.WalkerDeterminationoftheonsetofcrystallizationofN1-2-thiazolylsulfanilamidesulfathiazolebyUV-visandcalorimetryusinganautomatedreactionplatformsub-sequentcharacterizationofpolymorphicformsusingdispersiveRamanspectroscopySpectrochim.ActaA57917932001.doi10.1016S1386-14250100407-313.M.M.ParmarO.KhanL.SetonandJ.L.FordPolymorphselectionwithmorphologycontrolusingsolventsCryst.GrowthDes.716352007.doi10.1021cg070074n14.H.R.H.AliH.G.M.EdwardsandI.J.ScowenInsightintothermallyinducedsolid-statepolymorphictrans-formationofsulfathiazoleusingsimultaneousinsituRamanspectroscopyanddifferentialscanningcalorimetryJ.RamanSpectrosc.408872009.doi10.1002jrs.218915.G.J.KrugerandG.GafnerThecrystalstructuresofpolymorphsIandIIIofsulphathiazoleActaCrystallogr.B282721972.doi10.1107S056774087200218316.G.J.KrugerandG.GafnerThecrystalstructureofsulphathiazoleIIActaCrystallogr.B273261971.doi10.1107S056774087100217617.T.GelbrichD.S.HughesM.B.HursthouseandT.L.ThrelfallPackingsimilarityinpolymorphsof P.MacFhionnghaileetal.J.NearInfraredSpectrosc.215566201365sulfathiazoleCryst.Eng.Comm.1013282008.doi10.1039b807898j18.A.MunroeD.CrokerA.C.RasmusonandB.K.HodnettAnalysisofFIIcrystalsofsulfathiazoleepitaxialgrowthofFIIonFIVCryst.Eng.Comm.138312011.doi10.1039c0ce00221f19.S.N.CampbellRobertsA.C.WilliamsI.M.GrimseyandS.W.BoothQuantitativeanalysisofmannitolpolymorphs.X-raypowderdiffractometryexploringpreferredorientationeffectsJ.Pharm.Biomed.Anal.28611492002.doi10.1016S0731-70850200053-520.S.S.IyengarN.V.PhadnisandR.SuryanarayananQuantitativeanalysesofcomplexpharmaceuticalmixturesbytheRietveldmethodPowderDiffr.16202001.doi10.11541.133207621.V.UvarovandI.PopovDevelopmentandmetrologicalcharacterizationofquantitativeX-raydiffractionphaseanalysisforthemixturesofclopidogrelbisulphatepolymorphsJ.Pharm.Biomed.Anal.4646762008.doi10.1016j.jpba.2007.11.02622.W.DongC.GilmoreG.BarrC.DallmanN.FeederandS.TerryAquickmethodforthequantitativeanalysisofmixtures.1.PowderX-raydiffractionJ.Pharm.Sci.9722602008.doi10.1002jps.2114223.D.GironS.MonnierM.MutzP.PiechonT.BuserF.StowasserK.SchulzeandM.BellusComparisonofquantitativemethodsforanalysisofpolyphasicpharmaceuticalsJ.Therm.Anal.Calorim.897292007andrefstherein.doi10.1007s10973-006-7962-y24.A.D.PatelP.E.LunerandM.S.KemperLow-leveldeterminationofpolymorphcompositioninphysicalmixturesbynear-infraredreectancespectroscopyJ.Pharm.Sci.9033602001.doi10.10021520-60172001039033.0.CO2-U25.A.HskuldssonPredictionmethodsinscienceandtechnologyBasicTheoryVol.1.ColourscanWarsawPoland1996.26.P.GeladiandB.R.KowalskiPartialleast-squaresregressionatutorialAnal.Chim.Acta18511986.doi10.10160003-26708680028-927.M.OtsukaF.KatoandY.MatsudaDeterminationofindomethacinpolymorphiccontentsbychemometricnear-infraredspectroscopyandconventionalpowderX-raydiffractometryAnalyst126915782001.doi10.1039b103498g28.J.RantanenH.WikstrmF.E.RheaandL.S.TaylorImprovedunderstandingoffactorscontribut-ingtoquanticationofanhydratehydratepowdermixturesAppl.Spectrosc.5979422005.doi10.1366000370205441167029.P.E.LunerS.MajuruJ.J.SeyerandM.S.KemperQuantifyingcrystallineformcompositioninbinarypow-dermixturesusingnear-infraredreectancespectros-copyPharm.Dev.Technol.522312000.doi10.1081PDT-10010053830.K.PllnenA.HkkinenM.HuhtanenS.-P.ReinikainenM.KarjalainenJ.RantanenM.Louhi-KultanenandL.NystrmDRIFT-IRforquantitativecharacterizationofpolymorphiccompositionofsulfathiazoleAnal.Chim.Acta5441082005.doi10.1016j.aca.2005.02.07431.P.McArdleK.GilliganD.CunninghamR.DarkandM.MahonAmethodforthepredictionofthecrystalstructureofionicorganiccompoundsthecrystalstructuresofo-toluidiniumchlorideandbromideandpolymorphismofBicifadinehydrochlorideCryst..Eng.Comm.63032004.doi10.1039b407861f32.Z.NemetG.CsonkaKisG.PokolandA.DemeterQuantitativedeterminationoffamotidinepolymorphsX-raypowderdiffractometricandRamanspectrometricstudyJ.Pharm.Biomed.Anal.4923382009.doi10.1016j.jpba.2008.11.03333.N.BlagdenCrystalengineeringofpolymorphappearancethecaseofsulfathiazolePowderTechnol.121462001.doi10.1016S0032-59100100373-434.Y.HuA.ErxlebenA.G.RyderandP.McArdleQuantitativeanalysisofsulfathiazolepolymorphsinternarymixturesbyattenuatedtotalreectanceinfrarednear-infraredandRamanspectroscopyJ.Pharm.Biomed.Anal.5334122010.doi10.1016j.jpba.2010.05.00235.K.PllnenA.HkkinenS.-P.ReinikainenJ.RantanenM.KarjalainenM.Louhi-KultanenandL.NystrmIRspectroscopytogetherwithmultivariatedataanalysisasaprocessanalyticaltoolforin-linemonitoringofcrys-tallizationprocessandsolid-stateanalysisofcrystallineproductJ.Pharm.Biomed.Anal.3822752005.doi10.1016j.jpba.2005.01.00936.J.AaltonenJ.RantanenS.SiiriM.KarjalainenA.JrgensenN.LaitinenM.SavolainenP.SeitavuopioM.Louhi-KultanenandJ.YliruusiPolymorphscreeningusingnear-infraredspectroscopyAnal.Chem.7552672003.doi10.1021ac034205c37.M.S.DhanoaS.J.ListerR.SandersonandR.J.BarnesThelinkbetweenmultiplicativescattercorrectionMSCandstandardnormalvariateSNVtransformationsofNIRspectraJ.NearInfraredSpectrosc.2431994.doi10.1255jnirs.3038.I.S.HellandT.NaesandT.IsakssonRelatedversionsofthemultiplicativescattercorrectionmethodforpre-processingspectroscopicdataChemometr.Intell.Lab.Syst.292331995.doi10.10160169-74399500031-139.ICHHarmonizedguidelinep.Q2BinNoteforGuidanceonValidationofAnalyticalProceduresMethodologyp.1996.40.M.TiwariG.ChawlaandA.K.BansalQuanticationofolanzapinepolymorphsusingpowderX-raydiffractiontechniqueJ.Pharm.Biomed.Anal.433652007. 66LimitsofQuantitativeAnalysisofSulfathiazolePolymorphism41.Z.NemetA.DemeterandG.PokolQuantifyinglowlevelsofpolymorphicimpurityinclopidogrelbisulphatebyvibrationalspectroscopyandchemometricsJ.Pharm.Biomed.Anal.491322009.doi10.1016j.jpba.2008.09.04242.S.N.CampbellRobertsA.C.WilliamsI.M.GrimseyandS.W.BoothQuantitativeanalysisofmannitolpolymorphs.FT-RamanspectroscopyJ.Pharm.Biomed.Anal.28611352002.doi10.1016S0731-70850200059-643.M.R.AbuBakarZ.K.NagyC.D.RiellyandS.E.DannInvestigationoftheriddleofsulfathiazolepolymorphismInt.J.Pharm.41412862011.doi10.1016j.ijpharm.2011.05.004 JOURNALOFNEARINFRAREDSPECTROSCOPY67ISSN0967-0335IMPublicationsLLP2013doi10.1255jnirs.1035AllrightsreservedDuringthelasttenyearsnearinfraredNIRspectroscopyhasbeenincreasinglyappliedtosoilanalysis.1Asaresultofacontinuousdevelopmentofsensingtechnologiesawidevarietyofspectrometersisavailable.TheNIRinstrumentationcanbeassembledwithopticalcomponentsemployedforUV-visibleinstruments.BlancoandVillarroya2andPasquini3Comparingpredictiveabilitiesofthreevisible-nearinfraredspectrophotometersforsoilorganiccarbonandclaydeterminationMariaKnadelaBoStenbergbFanDengaAntonThomsenaandMogensHumlekrogGreveaaDepartmentofAgroecologyandEnvironmentFacultyofAgriculturalSciencesAarhusUniversityBlichersAll20POBox50DK-8830TjeleDenmark.E-mailmaria.knadelagrsci.dkbSwedishUniversityofAgriculturalSciencesDepartmentofSoilandEnvironmentSLUPOBox234SE-53223SkaraSwedenDuetoadvancesinopticaltechnologyawiderangeofspectrometersisavailable.Recentinterestsinsoilgloballibrariesandsensorfusionpresentsachallengewithrespecttocombiningdatafromdifferentinstrumentation.Littleresearchhoweverhasbeendoneonthecomparisonofvisible-nearinfraredvis-NIRspectrometersforsoilcharacterisation.Thereisaneedformoreworkontheeffectsofscanningstrategiesanduseofdifferentsoilinstrumentation.Wecomparedthreevis-NIRspectrometerswithvaryingresolutionsignal-to-noiseratiosandspectralrange.Theirperformancewasevaluatedbasedonspectracollectedfrom194DanishtopsoilsandusedtodeterminesoilorganiccarbonSOCandclaycontent.Scanningproceduresforthethreespectrophotometerswheredoneaccordingtouniformlaboratoryprotocols.SoilorganiccarbonandclaycalibrationswereperformedusingPLSregression.Onethirdofthedatasetwasusedasanindependenttestset.Arangeofspectralpreprocessingmethodswasappliedinsearchofmodelimprovement.ValidationforSOCcontentusinganindependentdatasetderivedfromallthreespectrophotometersprovidedvaluesofRMSEPbetween0.45and0.52r20.420.59andRPD1.21.4.ClaycontentwaspredictedwithahigherprecisionresultinginRMSEPvaluesbetween2.6and2.9r20.710.77andRPDvaluesintherangefrom2.2to2.5.NosubstantialdifferencesinthepredictionaccuracywerefoundforthethreespectrometersalthoughtherewasatendencythatinthetradeoffbetweennoiseandresolutionlownoisewasthemoreimportantforSOCandclaypredictions.Theapplicationofdifferentspectralpreprocessingproceduresdidnotgenerateimportantimprovementsofthecalibrationmodelseither.Additionallydatasimulationanalysisincludingresamplingtoacoarserresolutionandadditionofnoisewasperformed.Noorverylittleeffectofsamplingresolutionandadditionalnoiseontheperformanceofthespectrophotometerswasreported.Theresultsfromthisstudyshowedthataslongasstrictlaboratoryscanningprotocolswerefollowednosignicantdifferencesinconstituentdeterminationwerefounddespitedifferencesinspectralrangespectralresolutionspectralsamplingintervalsandsamplepresentationmethods.Thedifferencesinpredictiveabilitiesbetweenthespectrometersweremostlyduetodifferencesinspectralrange.Keywordsvis-NIRspectroscopypredictiveabilitysoilorganiccarbonSOCclayspectrophotometercomparisonIntroductionM.Knadeletal.J.NearInfraredSpectrosc.2167802013Received19April2012Revised12October2012Accepted22October2012Publication21January2013 68SOCandClayDeterminationUsingThreeVis-NIRSpectrometerspresentedreviewsonNIRinstrumentation.Themainadvan-tagesoftheNIRtechniquearespeedofanalysislittleornosamplepreparationandabilitytoperformmeasurementsintheeld.Inadditiontolaboratoryequipmentsmallportablespectrophotometerssuitableforinsitumeasurementsareavailable.Theyincludehand-heldinstrumentsorspectrom-etersthatcanbecarriedinabackpack.Mostrecentlymobilespectrometerplatformsdesignedforon-the-goeldsurveyhavegainedinterest.48Promisingresultsreportedfortheregionalnationalandgloballibraries914hasledtothedevelopmentofauniversalandstandardisedsoillibrary.Thecreationofalibraryhasbeenputforwardinordertoincreasetheefciencyofvis-NIRcharacterisation.1516Fortheconstructionofgloballibrariesespeciallythiscouldmeantheamalgamationofsensoryandanalyticaldataobtainedfromarangeofinstrumentsandmethods.Thesuccessofspectroscopicmodellingofsoilpropertiesishoweverdependentontheaccuracyandreliabilityofthereferencemethods.17Additionallysub-samplingerrorandtheuseoffundamentallydifferentreferencemethodscanpresentsourcesoferrorornoisetoNIRcalibrations.18Thusincludingtheresultsfromavarietyofanalyticalmethodsusedbydifferentlaboratoriesforcorrelationwithspectroscopicdatamaynotbesuccessfuloratallpossible.Thecalibrationmayproducehighvalidationerrorsbutaslongastheanalyticalmethodsarewellcorrelatedtoeachotherthepredictionsmaybeasgoodasifusingonesinglemethod.Moreoverthescanningprotocolsusedbytheindividuallaboratoriesandtheemployedinstrumentationmayvarysubstantially.AsreportedbyIgneetal.19differencesintechnologiesinternalinstrumentfeaturessamplepreparationandpresentationandthenumberofreplicateshavealargeimpactonthenalcalibrationmodels.Pimsteinetal.20dividedthepossiblefactorsaffectingreectancemeasurementintospectrometerandsamplingdomains.Forspectrometersspectralconfigurationdetectorperformanceopticalcharacteristicssurroundingfactorswarm-uptimeandcalibrationqualitywerenamedwhereasforsamplingfactorsincludedsamplehomogeneitygeometryofmeasurementilluminationsetupreferencemethodmeasurementconditionsandtheeffectsoriginatingfromtheoperatorinputwerelisted.Neverthelesslittleworkhasbeendoneontheeffectsofusingdifferentspectrophotometersandscanningmethodsonspectraqualityandtheprecisionofcalibrationmodels.Pimsteinetal.20comparedspectrometersofthesamemodelandvendorFieldSpecASDInc.BoulderCOUSAusingthreestandardprotocolsandusingthreedifferentmaterialsforinternalstandardssandglassgreyreferenceononly12soilsselectedfromanIsraelisoillibrary.Theyrecommendedusingastrictscanningprotocolandaninternalstandardfortheimprovementofspectralmeasurementsandtheirstabilitybyminimisinguncontrolledspectralvariations.Mouazenetal.21comparedfourcommerciallyavailablespectrophotometersfordifferentagriculturalmaterialsincludingsoilmoisturecontent.Thedifferencesbetweentheseinstrumentswerenotonlythewavelengthrangeandmeasurementresolutionbutalsothemeasurementprinciples.Despitethedifferencesinmeasurementprincipalsofthespectrophotometersuseda3001700nmdiodearraya3502500nmdiodearrayscanningmonochromatoranda4002500nmFouriertransformandascanningmonochromatorverysmalldifferencesfortheestimationofsoilmoisturewerefoundbetweenthem.Geetal.22comparedSOCcalibrationswithfourspectrophotometersAgriSpec352500nmNIRSystems65004002498nmLabSpec50003502500nmandaFieldSpecPro350nm2500nmontwosetsofsoilsamples.Apartfromcomparingdifferentinstrumentationdifferentprecisionsofscanningproceduresweretested.SignicantdifferenceswerefoundforcalibrationswhennostrictattentionwasgiventothemeasurementprocedureswhereasmoreconsistencyinsoilspectraandSOCcalibrationmodelsamongthesespectrophotometerswasobtainedbyhigherscanningcontrol.Igneetal.19focusedontheevaluationofspectralpretreatmentsfordifferentcalibrationmethodsofsoilconstituentsincludingSOCandclay.Measurementswereobtainedonair-dryandeld-moistsoilsusingfourdifferentspectrophotometersincludingportableFouriertransformIR4000400cm1bench-Fouriertransform-NIRbench-Fouriertransform-mid-IRandVeris3502225nm.Theyreportednosignicantdifferencesexistingamongpretreatmentsmethods.Whilecalibrationtransferbetweendifferentinstrumentshasbeensuccessful22stillmoreresearchisneededtoachievecomparabilitybetweeninstruments.11Moreoverduetothenumberofspectrophotometersavailabletheselectionofthesuitableinstrumentforaspecificapplicationcanbedifcult.21Thechoiceofinstrumentisrstofallapplication-dependentbutsoalsoistheissueofcost.Thequestionofmulti-functionalityandflexibilityisalsooftenaddressed.Thusacomparisonofpredictiveabilitiesofdifferenttypesofinstrumentswouldbeuseful.Theaimofthisstudywastocomparethepredictiveabilitiesofthreecommerciallyavailablevis-NIRspectrophotometerscommonlyusedforsoilspectroscopy.Twoofthespectrophotometerswerebasicallythesamehighprecisioninstrumentscoveringthefullvis-NIRrange3502500nmbutwithsomewhatdifferentresolutionandnoiselevels.Theirsamplingintervalswerenarrowandtheywereequippedwithtwodifferentbutcommonlyusedprobes.The3rdinstrumenthadashorterspectralrange3502200nmintermediateresolutionbutbroadersamplingintervals.TheirabilitiestopredictSOCandclayinrepresentativetopsoilsfromDenmarkweretested.MaterialandmethodsSoilsamplesOnehundredandninety-fourtopsoilsamplesFigure1wereselectedfromtheDanishsoildatabase.Soilsfromthenational M.Knadeletal.J.NearInfraredSpectrosc.216780201369databasewerecollectedduringaregionalsoilproleinvestiga-tionona7kmgridbetween1987and1989.23ThesampleswerechosentocoverarepresentativerangeofsoiltypesacrosstheentirecountryEastJutlandMiddleJutlandWestJutlandNorthJutlandBornholmDjurslandHimmerlandNorthZealandandThyandwereclassiedasGleysolPhaeozemCambisolLuvisolPodzolArenosolRegosolFluvisolAlisolHistosolAnthrosol.24ArangeoflandusesystemswerealsorepresentedcerealfieldcropspastureconiferousforestleguminousplantsfallowcruciferaeplantsBetavulgarisandpotatoes.Soilsamplesweredriedandsieveddownto2mmandwereanalysedforsoilorganiccarbonSOCandclay.Particlesizedistributionwasdeterminedbysievingandhydrometricmethods.25SOCwasdeterminedbycombustionusingaLECOinductionfurnaceCN-2000instrumentLECOCorporationStJosephMIUSA.BothSOCandclayarereportedin.SpectrophotometersThreedifferentspectrophotometerswereusedforscanningthesoilsamplesAbenchtopvis-NIRinstrumentLabSpec5100manufac-turedbyASDInc.BoulderCOUSAAportablevis-NIRinstrumentFieldSpecProFRalsomanufacturedbyASDInc.BoulderCOUSAAmobileshank-basedvis-NIRinstrumentinabenchtopmodeVerisTechnologiesKSUSA.Table1presentsaninstrumentoverviewincludingtheinternalfeaturesofthethreespectrometers.Spectralresolu-tionandsamplingintervalsarelisted.SpectralresolutionasdenedbyASDInc.isthefull-width-half-maximumFWHMoftheinstrumentresponsetoamonochromaticsourcewhereasspectralsamplingintervalisthespacingbetweensamplepointsinthespectrum.EventhoughbothASDInc.instrumentsemploythesamedetectorstheyareconfigureddifferently.TheFieldSpecisdesignedforuseundersolarilluminationanditsintegratedbreopticreducessignallossbecauseitiscontinuousfromthedetectortothetipwhereastheopticintheLabSpechastwojunctions.Oneisinternalattherearofthescramblerandtheotherisanexternalconnection.Approximately5ofthesignalislostateverycableinterfaceorjunction.ThereisFigure1.LocationofthesamplingpointsacrossDenmark.SpectrometermodelWavelengthrangeSpectralresolutionSamplingintervalOpticsDetectorsLabspec51003502500nm3nm700nm6nm14002100nm1.377nm3501050nm2nm10002500nmFibreopticFixedreectivegrating3501000nmmovinggratings10012500nm512elementSiphotodiodearray3501000nmtwoTEcooledInGaAsphotodiodes10011830nmand18312500nmFieldSpecProFR3502500nm3nm700nm10nm14002100nm1.377nm3501050nm2nm10002500nmFibreopticFixedreectivegrating3501000nmmovinggratings10012500nm512elementSiphotodiodearray3501000nmtwoTEcooledInGaAsphotodiodes10011830nmand18312500nmVeris3502200nm8nm6nm3501000nm5nm10732200nmFibreopticFixedreectivegrating3648elementlinearCCDarray3501050nm256InGaAslinearimagespectrometer9002200nmTable1.Instrumentationoverview. 70SOCandClayDeterminationUsingThreeVis-NIRSpectrometersalsoatrade-offbetweenresolutionandsignalstrength.Thusdifferencesinnoiselevelscanbeexpected.SamplepreparationandspectralmeasurementsAllspectralmeasurementsweremadeunderlabora-toryconditionsonairdried2mmsievedandhomogenisedsamples.Controlledscanningenvironmentsincludinganinstrumentspeciccheckproceduresamplepreparationandscanningprocedureadequatefortheindividualinstrumentwereassuredfollowingtherespectivelaboratoryspectropho-tometerprotocols.Tostabilisetheinstrumenttemperatureaminimumofhalfanhourwarmingwasallowedbeforescan-ningsessionsforallthreeinstruments.Astheinitialpurposeofthesemeasurementswasnotthecomparisonofthethreeinstrumentsdifferentscanningprotocolswerefollowedasdescribedbelow.LabSpec5100Thespectrophotometercoversthefullvisibleandnearinfraredrangebetween350nmand2500nm.Ithasthehighestresolu-tionofthethreeinstrumentsintheNIRregionTable1.ThesoilsweremeasuredusingHighIntensityMuglightmodel-A122106ASDInc.equippedwithasapphirewindowusinganASDIsamplingtrayadapterwithaquartzwindowapprox.10gofsoilhavinga110mm2spotdiameterASDInc..Theprobefeaturesabuilt-inlightsourceandactsasaworkstationsothatsamplescanbeplacedontopoftheprobe.Thesourceoflightisatungstenquartzhalogenlamp4W3.8Vwithbuilt-inDCcurrentstabilisercircuitry.Thecolourtemperatureofthehalogenbulbis2901K10K.TheLabSpechasaremovablebreoptic25fullconicalanglethatrequirestheuseofalaboratorystyleaccessory.Tworeplicatesofeachsamplewerescanned.Onespec-trumfromeachreplicatewascollectedbyIndicoPro6.0spec-trumacquisitionsoftwareASDInc.BoulderCOUSAandthenaveraged.ALabsphereSpectralonwww.labsphere.comproductsreflectance-standards-and-targetsreflectance-targetsspectralon-targets.aspxwhitereferencewasusedatthebeginningofeachscanningsessionandaftereveryfifthsample.Theinternaldarkcurrentwasacquiredauto-maticallybeforeeachmeasurementofthewhitereference.Themeasurementsofthesoilsampleswhitereferenceanddarkcurrentwereconguredasanaverageof50readingspercollectedspectrum.FieldSpecProFRFieldSpecProincludesthesamedetectorsasLabSpec5100butthecongurationisdifferentTable1andtheopticbreisintegratedwithoutjunctions.MeasurementsweremadeusingtheHighIntensityContactProbeASDInc.BoulderCOUSAequippedwithasapphirewindow.Ithasa100mm2spotsizeabuilt-inDCcurrentstabilisercircuitryandisequippedwithatungstenquartzhalogenlampwiththesamespecicationasforLabSpec.Thebreoptichasa25fullconicalangleandis1.5mlongdirectlyconnectedwiththedetectors.TheRS3WindowsinterfacesoftwarefromASDInc.wasusedfordataacquisition.Threereplicatespectraofeachsampleweretakenandaveragedrotatingthesamplecontaineraftereachmeasurement.ALabsphereSpectralonwhitereferencewasusedatthebeginningofeachscanningsessionandaftereveryfthsample.Theinternaldarkcurrentwasacquiredautomaticallybeforeeachmeasurementofthewhiterefer-ence.Themeasurementsofthesoilsampleswhitereferenceanddarkcurrentwereconguredasanaverageof30readingspercollectedspectrum.VerisTheVerisspectrophotometerispartofamobilesensorplatformdesignedmainlyforon-the-gomeasurementsintheeld.Inthisstudyitwasusedinbenchtopmodeinthelaboratory.Itcomprisestwodetectorsforspectraldatacollectionatarangebetween350nmand2200nmwith8nmspectralresolutionwhichisintermediatetothetwootherinstrumentsintheNIRregionbutlowerinthevisibleTable1.ThespectrometersusedwereanOceanOpticsUSB4000-vis-NIROceanOpticInstrumentsInc.DunedinFLUSAcharge-coupleddeviceCCDarray3501050nmandaHamamatsuC9914GBTG-CooledNIRIIHamamatsuShizuokaJapanInGaAslinearimagespectrometer9002200nm.Theinstrumentmakesmeasurementsthroughasapphirewindowwitha314mm2spotsizemountedonthebottomoftheshank.Aswiththeothertwospectrometersthisinstrumentusedatungstenhalogenbulb5V6.65W2700Ktoilluminatethesoilandabreopticof1.5metres.Inordertoacquirethedarkcurrentandthereferencespectratheshutterwouldfirstclosecompletelypreventinganylightcomingintothebreopticandthenmoveaknownreferencematerialinfrontoftheoptic.Thereferencemeasurementisusedtocompensatefordriftinthespectrometerandlightsource.SoilsampleswerepackedinaVerissampleholderapprox.1gofsoilandplacedagainstthefaceoftheshankwindowforscanning.Sampleswerescannedonlyoncewheretheoutputspectrumwasaresultofanaverageof100scans.Dataacqui-sitionandprocessingprogramswerecarriedoutusingtheVerisSpectrophotometerSoftwareV1.2NationalInstrumentsAustinTXUSA.SpectracomparisonSpectraldatausedintheanalysiscoveredtherangebetween450nmand2500nmfortheLabSpecandFieldSpecinstru-mentsandtherangebetween420nmand2155nmfortheVerisspectrophotometer.Inordertoeasethevisualcomparisonof M.Knadeletal.J.NearInfraredSpectrosc.216780201371absorptionfeaturesthespectrageneratedusingtheLabSpecandFielSpecwerereducedbyaveragetothesamesamplingintervalasfortheVerisspectrophotometerevery6nmat3501000nmandevery5nmat10732200nm.MultivariatedataanalysisBeforefurtherdataanalysistheverynoisyregionsneartheedgesofthespectrawereremoved.Thustheanalysisofspec-traldataincludedrangesof4252500nm4522500nmand4202158nmforLabSpecFieldSpecandVerisrespectively.PrincipalcomponentanalysisPCAanalysisPrincialcomponentanalysiswasperformedonallthreedatasetsseparatelyafterapplying1stderivativetransform.ThePCAwasusedtoexplorethedataandhowsamplesrelatetoeachotherwithinthespectraldataspacesascollectedbythethreespectrophotometers.Principalcomponentanalysisisananalysisappropriateforrevealingpatternsandinternalstructureofthedata.Itcanexplaintherelationshipsbetweensamplesandvariablesgivingageneraloverviewofthemaininformationcontentinthedatasetallowingfortheinterpretationofsamplegroup-ingssimilaritiesanddifferences.AprincipalcomponentPCisalinearrepresentationofvariationinthedata.EachPCexplainsamaximumamountoftheremaininginformationcontainedinthedata.ThustherstPCcontainsthegreatestamountofinformationfromthedatasetandeachsubsequentPCcontainslessinformationthanthepreviousone.AllPCswillalsobeorthogonaltoeachother.Partialleast-squaresPLSregressionDeterminationsofSOCandclayfromspectraldatawereperformedonapparentabsorbanceAlog1RwhereAabsorbanceRreectanceusingPLSregressionwithanon-lineariterativepartialleastsquaresNIPALSalgorithm26inUnscramblerX10.1softwareCamoASAOsloNorway.Partialleastsquaresregressionsonmeancantreddataweredevelopedusingacalibrationandatestsetforanindependentvalidation.Ineachcasesamplesweresplitintocalibration129samplesandvalidation65samplessets.ValidationdatasetsforSOCandclayweregeneratedbyrstsortingtheirvaluesandsecondbyselectingeverythirdsamplestartingfromnumbertwo.Theremainingsampleswereusedforcalibration.Thiswasdonetoassurefullrangecoverageofbothconstituentsinboththecalibrationandvali-dationdatasets.Fullcross-validationwasusedtodevelopthecalibrationmodelsforbothconstituentswiththethreespectrophotometerstondtheoptimumnumberofcompo-nentswithoutover-tting.Differentspectralpre-treatmentsweretestedinordertoimprovethecalibrationbasedonrawdatastandardnormalvariateSNVwithde-trending27afullmultiplicativescattercorrectionMSC28andtransformationstothefirstandsecondderivativeweregenerated.ForthederivativesaSavitzkyandGolaySG29smoothingwitha2ndorderpolynomialover15wavelengthswasappliedhoweverdifferentoptionsweretestedheretochoosethebestoneresultsnotshown.Thebesttreatmentforcross-validationwasconsideredtobetheoneresultinginamodelwiththelowestrootmeansquareerrorofcross-validationRMSECVthehighestR2forcalibrationtherawR-squareofthemodelthehighestR2cvforthevalidationdatasetadjustedR-squareshowinghowgoodatcanbeexpectedforfuturepredictionsandthehighestRPDratioofstandarderrorofcross-valida-tiontostandarddeviation.17DatasimulationAsshowninTable1thethreespectrophotometerscomparedinthisstudydifferinspectralrangeandspectralresolutionTable1.Inordertotesttheeffectsofsamplingresolutionandrandomnoiseontheperformanceofthespectrophotometersdatasimulationincludingresamplingtoacoarserresolutionandadditionofnoisewasperformed.ResamplingThesamplingresolutionoftheLabSpecspectrophotometerwasadjustedtosimulatethecoarserresolutionofFieldSpecabove1000nmbyperformingamovingaveragetransformation.Inthistransformationdataaresmoothedaccordingtoachosensizeofsegmentswhichspecifyhowmanyadja-centcolumnsshouldbeusedtocomputetheaveragevalue.Thiswayforeachpointofthecurveamovingaverageiscomputedastheaverageoverasegmentencompassingthecurrentpoint.Theindividualvaluesarereplacedbythecorre-spondingmovingaverages.ResamplingofLabSpecdatawereperformedforthespectralrangeabove1000nmonlyatthreelevelsusingthreefiveandsevenpointsineachsegmentrespectively.InadditionthesamplingresolutionoftheLasbSpecandFieldSpecspectrometerswasre-sampledtosimulatethecoarserresolutionoftheVerisspectrometer.Againamovingaveragetransformationatthreelevelswasusedwithvesevenandninepointsinthespectralregionbetween350nmand1000nmandwiththreefiveandsevenpointsabove1000nm.ThespectralrangeoftheLabSpecandFieldSpecspectrophotometerswasalsoadjustedtotheVerisrange4202158nm.Inthelaststepsmoothedspectrawerereducedbysixandvebelowandabove1000nmrespec-tivelytoobtainthesamenumberofvariablesastheVerisspectrophotometer. 72SOCandClayDeterminationUsingThreeVis-NIRSpectrometersNoisesimulationInordertocomparesignal-to-noiseratiosfromLabSpecandFieldSpecspectrophotometers30spectraofaLabsphereSpectralonwitha10scanaveragewerecollectedfrombothinstruments.ThelevelofnoiseforLabSpecandFieldSpecspectrawasintheareaofCVcoefcientofvariationequalto0.5and0.3respectively.TheobtainedspectraoftheLabsphereSpectralonwereconvertedtoabsorb-ance.Signal-to-noiseratiowascalculatedasstandarddeviation.TheaveragestandarddeviationwashigherfromtheLabSpec0.00025thanfromtheFieldSpecspectra0.00015.InordertoincreasetheaveragevalueofthestandarddeviationofFieldSpecspectra0.00015ofaddednoisewasneeded.TosimulatetheeffectsofrandomnoisepresentinthespectraldatafromLabSpecspectrophotom-eterdifferentlevelsofnoisewereaddedtotheFieldSpecspectra.Asastartingpoint0.00015ofnoisewasadded.Furthermore0.0015and0.015ofnoisewereaddedtotestwhennoiseadditionhasasignicanteffectonspectralcalibration.AfterresamplingornoiseadditionthePLSmodelsforSOCandclaywererunagainandcomparedwithcalibrationresultsbasedontheoriginaldatasets.ResultsanddiscussionSelecteddescriptivestatisticsofSOCandclaydistributionsarepresentedbybox-whiskerplotsinFigure2andinTable2.SpectracomparisonMeanreectancespectraofthethreespectrophotometerswereplottedFigure3afterapplyingtheSNVtransform.Byandlargetheshapesoftheaveragespectraofthethreeinstrumentsaresimilar.InparticularthespectrafromtheASDILabSpecandFieldSpecaresimilarascouldbeexpectedgiventheirtechnicalsimilarities.Clearlyvisibledifferencesareneverthelesspresent.Thedifferencesarevisibleespe-ciallyatahigherresolutionwiththemeanspectrabeingmoresimilarinthesmallerscalefeaturesFigure3.ThemaincharacteristicofthereflectanceabsorptionfeaturesoccurringintheNIRregionforthethreespectraarelocatedat1400nmand1900nmandareassociatedwithwaterbands.Thesestrongabsorptionbandsarecausedbyovertonesandcombinationsofthethreevibrationfunda-mentalsofH2O.ItistheovertonesofOHsymmetricandaymmetricstretchingthatabsorbnear1400nmandthecombinationsofHOHbendingandOHstretchingthatabsorbnear1900nmandareinvolvedinclaymineralabsorptionfeatures.30Inordertoshowsmall-scaledifferencesinresolutionandnoisespectraofthesamplewiththehighestorganicmattercontentSOC3.9wereplotted.Figure4ashowsaspec-trumofthesamesampleintheregionholdinginformationonsoilorganicmatter1695nm1800nmscannedbyallinstruments.Inaccordancewithwhatcouldbeexpectedfromthelowerresolutionandtheunbrokenbre-opticconnectionoftheFieldSpecthisspectrumisthesmoothestofthethree.ThespectrumobtainedbytheLabSpecmodelwhichhastheFigure2.Box-whiskerplotsaforsoilorganiccarbonandbforclay.Thebottomandtopoftheboxrepresentthe25thand75thpercentilerespectively.Thebandnearthemiddleoftheboxisthemedian.Theendsofthewhiskersrepresentthe5thandthe95thpercentile.Thedotrepresentsoutliers.ConstituentDatasetMinimumMaximumMeanMedianSDSOCCalibration0.643.891.71.510.68Test0.753.821.721.510.70clayCalibration2.551.79.538.26.39Test2.524.99.438.35.49SDisstandarddeviationSOCissoilorganiccarbon.Table2.Generalstatisticsofsoilorganiccarbonandclayinthecalibrationandtestsets.ab M.Knadeletal.J.NearInfraredSpectrosc.216780201373highestresolutionandabrewithtwojunctionsshowsthehighestdegreeofsmall-scalenoise.AcomparisonofLabSpecandFieldSpecspectrabetween20202500nmareshowninFigure4b.Thisregioncarriesinformationonmineralandorganicmattercontent.Thesamelarge-scaleabsorptionfeaturescanberecognisedforbothspectra.ThespectrumobtainedusingtheLabSpecinstrumentwasagainevidentlymorenoisythanthespectrafromtheFieldSpecinstrument.PrincipalcomponentanalysisInthePCAanalysessimilarpatternsrepresentingallthreespectrophotometerswereapparentforthefirstthreePCsFigure5.Therstthreecomponentsexplained9591and95ofthetotalvariationforLabSpecFieldSpecandVerisspectrophotometerrespectively.InordertoeasethegeneraloverviewofthemaininformationcontentinthedatasetsPCscoresweregroupedintothreeclassesaccordingtoclaycontent.Despitesomedifferencesamongthemeanspectraobtainedfromthethreespectrophotometersasdiscussedabovethepatternswithinthescoreplotswerenearlythesameindicatingthatbasicallythesameinforma-tionwasobtainedusingallthreespectrophotometers.WhencomparingPCscoresfromLabSpecandFieldSpecdatainparticularthemainpatternswithinthesampleswerealmostidentical.DuetothenatureofPCAwitheachcomponentasalinearrepresentationofthemaximumun-explainedvariationtherstthreecomponentswillfocusonlargerscalevaria-tion.26Thismeansthatmoresubtlebutimportantdifferencesmaystilloccur.TheclaygroupingsforthePC1vsPC2plotsindicatethatPC1containedsubstantialinformationrelatedtoclaycontent.SimilarclearpatternsforSOCcouldnotbefoundnotshown.Fewpotentialoutlierswererevealed.ThesewerehoweversampleswiththemostextremeclayandSOCcontent.Asfarasgeneralspectraldatastructuresareconcernednoimportantdifferencebetweenthespectrophotometerswereobserved.PartialleastsquaresregressionCross-validationresultsforSOCandclayareshowninTables3and4respectively.Theindependentvalidationresultsfromthebestmodelsbasedoncross-validationresultsareshowninFigure6.AspredictionresultswerenotsubstantiallyworsewhencomparedwiththeASDIinstrumentsthepotentialdrawbackofonlyonereplicatewiththeVeriswasapparentlycompensatedforbythelargerviewedareausedwiththisinstrumentandthecoarserspectralsamplingintervalwasnotdetrimentaleither.Figure3.StandardnormalvariateSNVtransformedmeanreectancespectrumobtainedwiththreeinstrumentsLabSpecFieldSpecandtheVeris.Figure4.aSpectrumofasamplewiththehighestsoilorganiccarboncontent3.9intherangebetween1695nmand1800nmgeneratedbyVerisLabSpecandFieldSpecspectrophotometers.bSpectrumofasamplewiththehighestsoilorganiccarboncontent3.9intherangebetween2200nmand2500nmgeneratedfromLabSpecandFieldSpecspectrophotometer.ba 74SOCandClayDeterminationUsingThreeVis-NIRSpectrometersThepredictiveabilityforSOCwaslowfromallthreespectrophotometers.Lowerpredictiveabilitieswererecordedfortheindependentvalidationincomparisonwithcross-validationresultsforbothASDIinstruments.InrelationtothestandarddeviationofthedatasettheRMSEPwasabout50higherthanexpectedfromacompilationofpreviouslypublisheddata.31ThefinalstatisticsfortheindependentvalidationshowthehighestpredictiveabilityforSOCfromtheFieldSpecdatausing1stderivativespectra.VerysimilarresultsweregeneratedfromtheVerisspectrometer.ModelsderivedfromtheLabSpecspectrahoweverperformedwithslightlylowerprecision.ClaycontentwasdeterminedwithahigherprecisionthanSOCandwithsomewhatlowerRMSEPthanexpectedfrompreviouslypublishedresults.31TheindependentvalidationshowedlowerpredictiveabilitiesforLabSpecandVeristhantheresultsfromcross-validation.HigherpredictiveabilitywasonceagainrecordedforthevalidationresultsbasedondataobtainedwiththeFieldSpecinstrument.ItwastheMSCtransformedspectrathatproducedthebestmodelforclaydeterminationfromthisinstrument.Theresultsoftheinde-pendentvalidationfortheLabSpecandVeriswerealmostidentical.ThebestmodelsweregeneratedwithMSCandSNVde-trendingtransformedspectraforLabSpecandVerisrespectively.Whenusingthepredictionstatisticsofthetwosoilconstit-uentstoevaluateandcomparethespectrometersperfor-manceitwastheFieldSpecinstrumentthatdeliveredthebestpredictionresultsforSOCandclay.ThedifferenceswereverysmallsoitisdifculttoconcludeiftheyareduetochanceortoslightlymorerobustcalibrationsduetolessnoiseintheFieldSpecspectra.InanycaseourdatadonotsupportthehypothesisthatahigherspectralresolutionaswiththeLabSpecinstrumentbringsadditionalspectralinformationaboutSOCandclaycontent.DespitethelowerspectralrangeandbroadersamplingintervalsoftheVerismeasurementspredictionresultsforSOCcontentweresimilartotheASDInc.instruments.Figure5.ScoreplotsfortherstthreeprincipalcomponentsinaprincipalcomponentanalysisPCAofspectraldatafromthreeinstrumentsforLabSpecFieldSpecandVeris. M.Knadeletal.J.NearInfraredSpectrosc.216780201375Thatisprobablyexplainedbythefactthatorganicmatterisspectrallyactivethroughtheentirevis-NIRspectrum.TheovertonesandcombinationbandsherearetheresultofstretchingandbendingofNHCHandCOgroups.Themostimportantbandsforsoilorganicmatterarelocatedaround17001800nm2050nmand22002400nmbutbandsnear11001200nmand1600nmcanalsobeofsignif-icance.3132TheregressioncoefcientsforSOCmodelsonSNVtransformedspectraareshowninFigure7a.Thegeneralfeaturesweresimilarforallthreespectrophotom-eters.MorenoisehoweverwasintroducedintheSOCmodelderivedfromVerismeasuremtnsasaresultofthehighernumberoffactorsusedinthiscalibrationsixfactorsinsteadoffourwiththeothertwoinstruments.TheimportantbandsforSOCcalibrationfromthethreemodelsarelocatedatwavebandsaround600nm800nm1400nm17001800nm2070nm1900nmandadditionallyaround2200-2300nmforASDIinstrumentsFigure7a.Themajormineraldiagnosticregionsforclaymineralsarelocatedbetween13001400m18001900nmand22002500nmandareholdinginformationofthemajorclaymineralsinDenmarksuchassmectitekaolinorillite.Thespectralfeaturesofclaymineralsareduetoovertonesandcombina-tionvibtrationsmodesofOHfunctionalgroupsaslatticewateroraspartofabsorbedwater.30Howeverthelackofa22002500nmregionintheanalysisdidnotcauseadecreaseintheperformanceoftheVerisclaycalibrations.Itseemsasiftheinteractionsbetweenwaterandhydroxylsandclaymineralsnear1400nmand1900nm33wereabletocapturethevariationinclaycontentsufcientlyontheirown.Regressioncoef-cientsfromthethreeclaymodelspresentdistinctwavebandsatboth1400nmand1900nmFigure7b.Additionalimpor-tantwavebandsobviousforASDIinstrumentsarelocatedbetween2200nmand2300nm.Theresultsofothersreportedintheliteratureontheeffectsofusingdifferentspectrophotometersandscanningmethodsonspectralqualityandtheprecisionofcalibrationmodelsaresimilartoours.IncomparisonnomajordifferencesfortheestimationofsoilmoisturecontentamongfourspectrophotometerswerefoundbyMouazenetal.19Notonlyhadthewavelengthrangeandmeasurementresolutionoftheusedinstrumentsdifferedasincaseofourstudybutalsothemeasurementprinciples.LabSpecFieldSpecVerisCross-validationlog1RRMSECV0.420.460.49R2CV0.620.560.49RPD1.61.51.4Factors10671stSGRMSECV0.440.430.50R2CV0.590.600.47RPD1.51.61.4Factors8942ndSGRMSECV0.470.470.45R2CV0.530.530.53RPD1.41.41.4Factors337MSCRMSECV0.440.480.51R2CV0.580.500.44RPD1.51.41.3Factors855SNVRMSECV0.500.500.50R2CV0.470.460.46RPD1.31.41.4Factors446log1RisabsorbancewhereRisreectance1stSGisthe1stSavitzkyGolayderivative2ndSGisthe2ndSavitzkyGolayderivativeMSCismultiplicativescattercorrectionSNVisstandardnormalvariateRMSECVisarootmeansquareerrorofcross-validationRPDisratioofstandarderrorofcross-validationtostandarddeviationResultshighlightedinboldaretheresultsfromthebestcalibrations.Table3.Cross-validationon129samplesresultsforsoilorganiccarbonforthethreespectrometers.LabSpecFieldSpecVerisCross-validationlog1RRMSECV2.32.82.5R2CV0.870.810.84RPD2.82.32.5Factors7791stSGRMSECV2.52.72.5R2CV0.850.830.85RPD2.52.42.5Factors3472ndSGRMSECV2.52.72.5R2CV0.850.820.85RPD2.52.42.5Factors115MSCRMSECV2.32.32.5R2CV0.870.870.85RPD2.82.82.5Factors454SNVRMSECV2.42.82.5R2CV0.850.810.84RPD2.72.32.5Factors343ForabbreviationsrefertoTable3.Table4.Cross-validationresultsforclayforthethreespectrometers. 76SOCandClayDeterminationUsingThreeVis-NIRSpectrometersFigure6.IndependentvalidationresultsofsoilorganiccarbonandclayfortheLabSpecFieldSpecandVerisspectrometers. M.Knadeletal.J.NearInfraredSpectrosc.216780201377Yetthenalresultsseemednotbeaffectedsignicantly.InthestudybyGeetal.22SOCcalibrationsusingtwoandthreedifferentvis-NIRspectrophotometerswithandwithoutscan-ningcontrolrespectivelywerepresented.TheapproachofusingtwospectrometerswithahigherscanningcontrolismorerelevanttothemethodologypresentedinourstudyandshowedstatisticallyverysimilarSOCcalibrationmodelsregardlessofspectrometertype.Similartoourndingsnomajordifferencesamongvariousspectralpretreatmentswerefoundintheliterature.IntheworkbyIgneetal.19nosignicantlydifferentresultsforSOCandclaycalibrationswereobtainedamongasmanyas18spectralpretreatmentsfordatafromfourvis-NIRandMIRspectrophotometers.Vasquesetal.34alsoreportedthatonlyasmallgaininpredictiveaccuracywasachievedfromusingdifferentspectralpretreatmentsforSOCmodellingusingaPLSregressionmethodfor554soilsintheSantaFeRiverWatershedinFlorida.Theyfoundthatpreprocessingtransfor-mationsweremoreeffectiveforotherparametricmultivariatetechniques.Figure7.RegressioncoefcientsfromPLSregressionforStandardNormalVariatetransformeddataforsoilorganiccarbonaandclayb.ThenumberoffactorsusedinPLSforsoilorganiccarbondeterminationLabSpec4FieldSpec4andVeris6.FactorsusedinPLSforclaydeterminationLabSpec4FieldSpec4andVeris3.LabSpec4252500nmFieldSpec4522500nmVeris4202158nmLabspec4202158nmFieldSpec4202158nmRMSECV0.420.430.440.470.47R2CV0.620.600.580.530.53RPD1.61.61.41.41.4Factors109757ForabbreviationsrefertoTable3.LabSpec4252500nmFieldSpec4522500nmVeris4202158nmLabspec4202158nmFieldSpec4202158nmRMSECV2.32.32.52.32.6R2CV0.870.870.840.870.84RPD2.82.82.52.82.5Factors45355ForabbreviationsrefertoTable3.Table6.Thebestcross-validationresultsforclayfromtheLabSpecandFieldSpecoriginalspectraldatacoveringthespectralrangefrom425nmto2500nmandafterrangeadjustmentsfrom420nmto2158nmtomatchtheVerisspectralrange.Table5.Thebestcross-validationresultsforsoilorganiccarbonfromtheLabSpecandFieldSpecoriginalspectraldatacoveringthespectralrangefrom425nmto2500nmandafterrangeadjustmentsfrom420nmto2158nmtomatchtheVerisspectralrange.ab 78SOCandClayDeterminationUsingThreeVis-NIRSpectrometersDatasimulationResamplingNoeffectofresamplingLabSpecspectratosimulateFieldSpecresolutionwasfoundregardlessofthesizeofasmoothingsegmentinmovingaveragetransformation.Calibrationandvalidationmodelsshowedexactlythesameprecisionasmodelsbasedonoriginaldata.SimilarlynoeffectwasrecordedwhenadjustingLabSpecandFieldSpecresolutiontoVerisinternalsettings.AsresamplingLabSpecspectratoFieldSpecresolutiondidnotdegradecalibrationmodelsadditionaldatasimulationanalysisincludingresamplingtothepointwhenitdoesmakeadifferencewastested.Stillusingevenasmanyas100segmentsinmovingaveragetransformationdidnotaffectthecalibrationandvalidationofSOCandclay.Adegradedeffectoncalibrationmodelswasfirstvisibleafternarrowingthespectralrangedownto4202158nm.Tables5and6containcross-validationresultsfromthebestmodelsforSOCandclaycalibrationsrespectivelygeneratedontheoriginaldataandafternarrowingspec-tralrangetotheVerisrange.AllcalibrationmodelsfromLabSpecandFieldspecdataperformedworseafteradjust-mentsoftherange.TheresultsfromSOCmodelsafternarrowingthespectralrangewerenearlyidenticalforbothLabSpecandFieldSpecbeingofaslightlylowerpredictiveabilitythanfromVerisdataTable5.ThemostnoticeabledifferencewasthattheoptimumnumberofcomponentswasreducedtothelevelofVeriswhennarrowingthespec-tralrange.SlightlybetterresultsfromclaycalibrationwereobtainedonLabSpecthanFieldSpecdataTable6.Thebestcross-validationresultforclaygeneratedfromtheFieldSpecwithashorterrangeperformednearlyiden-ticallytotheVeris.Resultsfromtheindependentvalida-tionforbothSOCandclayTable7confirmeddegradedeffectsofshorterspectralrangeonthepredictiveabilityoftheASDinstrumentsandwiththeshorterrangediffer-encesbetweenallthreeinstrumentswereverysmallandrandom.NoisesimulationTheadditionofnoisetoFieldSpecspectraldataatdifferentlevels0.000150.0015and0.015tosimulateLabSpecsignal-to-noiseratiodidnotaffectthePLScalibrationsofSOCandclay.Nosignicantdifferenceinthecalibrationvalidationstatisticsrecordeddifferenceswereatthelevelofthefourthdigitnordifferencesinthenumberoffactorswerefoundafternoiseaddition.Thereforetheseresultsarenotshown.AhigherlevelofnoisepresentinLabSpecspectraldatawereremovedwiththeapplicationofspectralpretreatmentsandthushadnoeffectonthecalibrationresults.ConclusionsLaboratorymeasurementsonsoilsamplesundercontrolledlaboratoryconditionsusingthreecommerciallyavailablespectrophotometerswerecompared.Nosubstantialdiffer-encesbetweenpredictedSOCandclaycontentwerereported.Thoseminordifferencesincalibrationresultsaredifculttoexplainbutmightberelatedtodifferentscanningmethodsorsamplepresentation.CalibrationmodelsusingtheVerisinstrumentwhichoffersareducedspectralrangeweresurprisinglysimilartothemodelsobtainedbythetwoASDIspectrophotometers.InstrumentswithahigherresolutionorsampleintervaldidnotperformsubstantiallybetterforSOCandclaypredictionsthaninstrumentswithalowerresolution.Theresultsfromdatasimulationindicatethatdifferencesamongthepredictiveabilitiesofthespectrophotometerswererandomormightbedependentonthesamplepresentationandnotaffectedbynoiselevelorsamplingresolution.Theonlyfactorthathadsomeeffectwasspectralrange.Weconcludethatforthemeasurementsperformedincontrolledenvironmentscarefullyfollowingthelaboratoryprotocolsneitherthespectrophotometertypenorthescan-ningproceduresignicantlyaffectedthepredictiveabilitiesofthenalmodels.LabSpec4252500nmFieldSpec4522500nmVeris4202158nmLabspec4202158nmFieldSpec4202158nmSOCRMSEP0.520.450.550.580.58r20.490.590.420.350.36RPD1.31.41.21.21.2ClayRMSEP2.92.62.93.12.9r20.710.770.700.680.71RPD2.22.52.22.12.2RMSEPisaroommeansquareerrorofpredictionRPDisratioofstandarderrorofpredictiontostandarddeviationSOCissoilorganiccarbon.Table7.ResultsfromtheindependentvalidationfromfromtheLabSpecandFieldSpecoriginalspectraldatacoveringthespectralrangefrom425nmto2500nmandafterrangeadjustmentsfrom420nmto2158nmtomatchtheVerisspectralrange. M.Knadeletal.J.NearInfraredSpectrosc.216780201379Infutureworkcalibrationtransferbetweentheinstrumentswillbeinvestigated.AcknowledgementFinancialsupportforthisworkcamefromtheHOBEresearchcentreforhydrologyandtheSINKSproject.References1.V.Bellon-MaurelandA.McBratneyNear-infraredNIRandmid-infraredMIRspectroscopictech-niquesforassessingtheamountofcarbonstockinsoilscriticalreviewandresearchperspectivesSoilBiol.Biochem.4313982011.doi10.1016j.soil-bio.2011.02.0192.M.BlancoandI.VillarroyaNIRspectroscopyArapid-responseanalyticaltoolTrendsAnal.Chem.212402002.doi10.1016S0165-99360200404-13.C.PasquiniNearinfraredspectroscopyfundamen-talspracticalaspectsandanalyticalapplicationsJ.Braz.Chem.Soc.141982003.doi10.1590S0103-505320030002000064.M.KnadelA.ThomsenandM.H.GreveMultisensoron-the-gomappingofsoilorganiccarboncontentSoilsSci.Soc.Am.J.7517992011.doi10.2136sssaj2010.04525.X.HuangW.SenthilkurnarS.KravchenkoA.ThelenandJ.G.QiTotalcarbonmappinginglacialtillsoilsusingnear-infraredspectroscopylandsatimageryandtopographicalinformationGeoderma141342007.doi10.1016j.geoderma.2007.04.0236.C.D.ChristyReal-timemeasurementofsoilattributesusingon-the-gonearinfraredreectancespectroscopyComput.Electron.Agric.61102008.doi10.1016j.com-pag.2007.02.0107.A.M.MouazenM.R.MalekiJ.DeBaerdemaekerandH.RamonOn-linemeasurementofsomeselectedsoilpropertiesusingaVIS-NIRsensorSoilTillageRes.93132007.doi10.1016j.still.2006.03.0098.R.S.BricklemyerandD.J.BrownOn-the-goVisNIRpotentialandlimitationsformappingsoilclayandorganiccarbonComput.Electron.Agric.702092010.doi10.1016j.compag.2009.10.0069.K.D.ShepherdandM.G.WalshDevelopmentofreectancespectrallibrariesforcharacterizationofsoilpropertiesSoilsSci.Soc.Am.J.669882002.doi10.2136sssaj2002.098810.D.J.BrownK.D.ShepherdM.G.WalshM.D.MaysandT.G.ReinschGlobalsoilcharacterizationwithVNIRdiffusereectancespectroscopyGeoderma1322732006.doi10.1016j.geoderma.2005.04.02511.L.CcillonB.G.BarthesC.GomezD.ErtlenV.GenotM.HeddeA.StevensandJ.J.BrunAssessmentandmonitoringofsoilqualityusingnear-infraredreectancespectroscopyNIRSEur.J.SoilSci.607702009.doi10.1111j.1365-2389.2009.01178.x12.H.BellinasoJ.A.M.DematteandS.A.RomeiroSoilSpectralLibraryanditsuseinsoilclassicationRev.Bras.Cienc.Solo348612010.doi10.1590S0100-0683201000030002713.V.GenotG.ColinetL.BockD.VanvyveY.ReusenandP.DardenneNearinfraredreectancespectroscopyforestimatingsoilcharacteristicsvaluableinthediagnosisofsoilfertilityJ.NearInfraredSpectrosc.191172011.doi10.1255jnirs.92314M.KnadelF.DengA.ThomseandM.GreveDevelopmentofaDanishnationalVis-NIRsoilspectrallibraryforsoilorganiccarbondeterminationinDigitalSoilAssessmentsandBeyond.CRCPresspp.4034082012.doi10.1201b12728-7915.R.A.A.ViscarraRosselWorkinggrouponproximalsoilsensingWG-PSSPedometron.25272008.16.J.B.SankeyD.J.BrownM.L.BernardandR.L.LawrenceComparinglocalvs.globalvisibleandnear-infraredVisNIRdiffusereectancespectroscopyDRScalibrationsforthepredictionofsoilclayorganicCandinorganicCGeoderma1481492008.doi10.1016j.geoderma.2008.09.01917.P.C.WilliamsandK.H.NorrisImplementationofnear-infraredtechnologyinNear-InfraredTechnologyintheAgriculturalandFoodIndustry2ndEdnEdbyP.C.WilliamsandK.Norris.AmericanAssociationofCerealChemistsStPaulMNUSAp.1452001.18.D.F.MalleyP.D.MartinandE.Ben-DorApplicationinsoilanalysisinNearInfraredSpectroscopyinAgriculture.EdbyA.RobertsJ.WorkmanJrandJ.B.ReevesIII.Agronomy44Am.Soc.of.AgronInc.MadisonWisconsinUSApp.7297842004.19.B.IgneJ.B.ReevesG.McCartyW.D.HivelyE.LundandC.R.HurburghEvaluationofspectralpre-treatmentspartialleastsquaresleastsquaressupportvectormachinesandlocallyweightedregressionforquantitativespectroscopicanalysisofsoilsJ.NearInfraredSpectrosc.181672010.doi10.1255jnirs.88320.A.PimsteinG.NotescoandE.Ben-DorPerformanceofthreeidenticalspectrometersinretrievingsoilreec-tanceunderlaboratoryconditionsSoilsSci.Soc.Am.J.757462011.doi10.2136sssaj2010.017421.A.M.MouazenW.SaeysJ.XingJ.DeBaerdemaekerandH.RamonNearinfraredspectroscopyforagri-culturalmaterialsaninstrumentcomparisonJ.NearInfraredSpectrosc.13872005.doi10.1255jnirs.46122.Y.GeC.L.S.MorganS.GrunwaldD.J.BrownandD.V.SarkhotComparisonofsoilreectancespectraandcalibrationmodelsobtainedusingmultiplespectrom- 80SOCandClayDeterminationUsingThreeVis-NIRSpectrometersetersGeoderma1612022011.doi10.1016j.geo-derma.2010.12.02023.H.B.MadsenA.H.NrrandK.A.HolstTheDanishsoilclassicationinAtlasofDenmarkI3.ReitzelCopenhagenDenmark1992.24.L.KroghandM.H.GreveEvaluationofWorldreferencebaseforsoilresourcesandFAOsoilmapoftheWorldusingnationwidegridsoildatafromDenmarkSoilUseManag.151571999.25.G.W.GeeandJ.W.BauderPhysicalandchemicalmethodsinMethodsofSoilAnalysisPart1EdbyKluteSoilSci.Soc.Am.andAm.Soc.Agron.MadisonWIUSApp.3834121986.26.H.MartensandT.NsMultivariateCalibration.JohnWileySonsLtdChichesterUK1989.27.R.J.BarnesM.S.DhanoaandS.J.ListerStandardnormalvariatetransformationandde-trendingofnear-infrareddiffusereectancespectraAppl.Spectrosc.437721989.doi10.1366000370289420220128.H.MartensS.A.JensensandP.GeladiMultivariatelinearitytransformationsfornearinfraredreectancespectroscopyinProc.NordicSymp.AppliedStatisticsEdbyStokklandForlagStavangerNorwayp.2051983.29.A.SavitzkyandM.GolaySmoothinganddifferentiationofdatabysimpliedleastsquaresproceduresAnal.Chem.3616271964.doi10.1021ac60214a04730.R.N.ClarkSpectroscopyofrocksandmineralsandprinciplesofspectroscopyinRemoteSensingfortheEarthSciences.ManualofRemoteSensing.EdbyA.N.Rencz.JohnWileySonsLtdChichesterUKpp.3581999.31.B.StenbergR.A.ViscarraRosselA.M.MouazenandJ.WetterlindVisibleandnearinfraredspectroscopyinsoilscienceAdv.Agron.1071632010.doi10.1016S0065-21131007005-732.B.StenbergEffectsofsoilsamplepretreatmentsandstandardisedrewettingasinteractedwithsandclassesonvis-NIRpredictionsofclayandsoilorganiccarbonGeoderma158152010.doi10.1016j.geo-derma.2010.04.00833.R.N.ClarkT.V.V.KingM.KlejwaandA.SwayzeHighspectralresolutionreectancespectrosocpyofmin-eralsJ.Geophys.Res.95126531990.doi10.1029JB095iB08p1265334.G.M.VasquesS.GrunwaldandJ.O.SickmanComparisonofmultivariatemethodsforinferentialmodelingofsoilcarbonusingvisiblenear-infraredspectraGeoderma146142008.doi10.1016j.geo-derma.2008.04.007 JOURNALOFNEARINFRAREDSPECTROSCOPYJNIRSJournalofNearInfraredSpectroscopyaimstopublishoriginalresearchpapersshortcommunicationsreviewarti-clesandlettersconcernedwithnearinfraredspectroscopyandtechnologyitsapplicationnewinstrumentationandtheuseofchemometricanddatahandlingtechniqueswithinNIR.JNIRSwillacceptcontributionsfromallareasofapplicationswherenearinfraredspectroscopyisinusesuchasAgricultureChemicalIndustryFoodLifeSciencesProcessControlPharmaceuticalsTextilesPolymersWoodTopicsforpapersmayincludeChemometricsCalibrationsDiffuseReectionNIRImagingOn-LineUseFibreOpticsSamplingSpectroscopyInstrumentationRemoteSensingTypesofarticleAllarticlesshouldbewritteninEnglishandwillbesubjecttoarefereeingprocess.LettersmaybeacceptedforpublicationfollowingreviewbytheEditororaSubjectRegionalEditoralone.Originalresearchpapersshouldreportindetailtheresultsoftheauthorssworkinaparticularareaofresearch.Thisworkshouldhavebeencompletedorasubstantialandidenti-ablepartofitshouldhavebeencompleted.Shortcommunicationscanbeusedtoreportsignificantresultsfromincompleteworkandtoreportcompletedworkwhichdoesnotrequirethefullcoverageofaresearchpaper.Shortcommunicationsshouldberestrictedto1500wordsinlength.ReviewarticleswillnormallybecommissionedbytheEditororoneoftheRegionalorSubjectEditors.HowevershouldyoubeinterestedinreviewingaparticularareayouarewelcometodiscussthiswiththeEditororyourlocalRegionalorSubjectEditor.Reviewarticlesshouldprovideacomprehen-sivecoverageoftheareareviewedwithextensivereferencestopublishedwork.Letterscaneitherexpresstheviewoftheauthoronapartic-ularaspectofNIRspectroscopyandtechnologythattheywishtocommentonorcanverybrieyreportaninterestingaspectoftheauthorsworkperhapsdetailsofacalibrationorasamplingprocedure.ThemanuscriptForthefastestreviewandpublicationpleasesubmityourpaperonlineatwww.impublications.comauthors.Ifyoupreferanelectronicversionmaybesentbye-mailoronCD-ROMtotheEditor-in-ChiefthepublishersoraSubjectRegionalEditor.InstructionsforauthorsAimsandscopeDidyouknowJNIRSistheonlypeerreviewjournaldedicatedtoNIRspectroscopy.ThereareNOpagecharges.AllpapersarereviewedbyacknowledgedexpertsinNIRtechnology.Youareensuredthatwhenyourpaperhasbeenaccepteditwillbepublishedrapidly.JNIRSistheglobalaudienceforyourpaperonanyaspectofNIRspectroscopytechnologyoritsapplication.Youcansubmityourpaperonlineatwww.impublications.comauthors. Itisimportanttorememberthatthemanuscriptisgoingtobereadbyreferees.Pleaseassistthembyobservingthefollowingspecication.UseTimesorTimesNewRomanwithafontsizeofatleast12.Themanuscriptshouldbedoublelinespacedwithgenerousmargins.Pleasenumberthepagesandusethelinenumberoptionifyourwordprocessorhasit.JNIRSiscomputertypesetandauthorsshouldincludeacomputer-readableversionoftheirmanuscriptthereisnoneedforanyhardcopy.Thisnotonlyspeedsupthepublicationofthearticlebutalsoreducesthepossibilityforerrorsandreducestheproofreadingburdenforauthorandpublisheralike.WecanreadmostcommonPCwordprocessorles.WecanalsoacceptillustrationsincomputerreadableformsuchasPDFTIFFPCXorEPSformat.Pleaseensurethatthehighestqualitysettingsareusedfortheircreationandthatanyfontsareembeddedorconvertedtooutlines.Bitmapsshouldbeatleast300dpifortoneimagesandatleast1200dpiforlinework.PleaseavoidJPGsbecausetheyarealossyformateachtimeyousavealeasaJPGitlosesquality.Ifyoudonotsavewiththehighestqualitysettingsyourgurewillbeveryblurry.SymbolsmathematicsandformulaeThisisanareawherecareshouldbeexercised.YoushoulduseSIunitswheneverpossible.Pleaseusethefollowingconven-tionsallsymbolswhicharevariablesshouldbeitalicisedallconstantsshouldbenormalRomantextandvectorsandmatricesshouldbebold.ComplexformulaecanbepreparedinequationeditorsideallyMathType.PleasenotethatwedonothavethefacilitytoprocessteX.Howeverpleaseusetheequationeditoronlyfordisplayequationsthoseappearingontheirownlineandnotforitaliccharactersetc.Multi-lineequationscannotbeincludedwithinthetextowsopleaserewritetheseassingle-lineequationsorincludethemasdisplayedequationsontheirownline.Theexceptiontothisischaractersrequiringmodifyingmarkssuchashatorbar.UnitsUnitsofmeasurementshouldbeseparatedfromthenumberbyaspace.Pleaseavoidtheuseofthesolidusorperanduseasuperscript1i.e.5mgL1not5mgL.ReferencesReferencesshouldbecitedinthetextbyasuperscriptArabicnumberi.e.spectroscopy1andlistedattheendofthearticle.PleasegivefullbibliographicalinformationincludingthetitleofallpapersandbookchaptersasintheexamplesbelowForeditedbooks1.G.D.BattenA.B.BlakeneyandC.R.BlattSamplepreparationproceduresforuseinnearinfraredanalysisoffruitandvegetablecropsinNearInfraredSpectroscopyTheFutureWavesEdbyA.M.C.DaviesandP.Williams.NIRPublicationsChichesterp.1071996.Forjournals2.T.NsK.KvaalT.IsakssonandC.MillerArticialneuralnetworksinmultivariatecalibrationJ.NearInfraredSpectrosc.111993.Forauthoredbooks3.B.OsborneT.FearnandP.H.HindlePracticalNIRSpectroscopywithApplicationsinFoodandBeverageAnalysis.LongmanScienticandTechnicalHarlowUK1993.Itistheresponsibilityoftheauthortoensurethatpermis-sionisobtainedfromthecopyrightholdertoreproducecopy-rightmaterial.Thecopyrightholderisalmostalwaysthepublisherofthework.FormsforthispurposeareavailablefromtheEditorandfromIMPublications.CopiesofanysignedPermissionFormsshouldaccompanythemanuscriptandthenecessarywordingofacknowledgementrequestedbythecopyrightholdershouldbeincludedinthecaption.AbstractAshortabstractnotexceeding350wordsshouldbeincluded.Thisshoulddescribethescopeofthearticleandhighlightanysignicantareas.KeywordsUpto10Keywordsshouldbeincluded.IllustrationsIllustrationsmustbeproducedreadyforpublicationandmeetingourguidelinessetoutabove.Labelsontheaxesshouldbeclearandofsufcientsizethattheywillbelegiblewhenreducedtotonthepagei.e.afontsizeofatleast8ptwhenproducedatonecolumnwidth.Thisrequiresafontsizeofatleast20ptifthegureisdrawnfullsizeonA4paper.Notethatordinateandabscissascalesshouldbeequalforequalvalues.Refereesexpecttoseesquareplotsofresultsofforinstancepredictedversusreferencemethod.PleaseuseHelveticaorArialfont.Theamountoftextontheillustrationshouldbekepttoaminimum.Keysfordifferentlinestylesetc.shouldbegiveninthecaptionnotontheillustration.IllustrationsshouldbenumberedinthetextwithArabicnumeralsasFigure1Figure4and5etc.theyshouldbeinaseparatesequencefromanytables.Captionsshouldbeprovidedforallguresandtheseshouldappearattheendofthemanuscript.Ifembeddedinthewordprocessorleeachillustrationshouldbeprovidedonaseparatepageaslargeaspossible.PleaselabeleachpagewiththeFigurenumberauthorsandtitleofthepaper.Photographscanbepublishedbuttheymustbeofgoodphotographicqualityandsuppliedashighqualityimagesataminimumof300dpi.Colourillustrationscanonlybeincludediftheauthormeetstheextracostoftheirproduction.Howeveranyillustrationscanbereproducedincolourintheonlineeditionatnocostandinblack-and-whiteinprint.Ifyouinclude colourguresthiswillhappenautomaticallysoyoumustinformtheEditorifyouwishtopayforcolourreproduction.TablesTablesshouldbenumberedinaseparatesequencefromtheillustrationsasTable1etc.PleaseconsiderthesizeofthepagewhencompilingyourTable.Tablesshouldnotbesplitwidth-wiseacrosspagesalthoughiftheyexceedapageinlengththeycanbecontinuedonsubsequentpages.Tablesshouldbeprovidedonseparatepagesandnotincludedinthetext.EachTableshouldhaveacaption.Aswithillustrationspleasenumbereachsheetandincludetheauthorsandtitleofthepaper.CopyrightAuthorsarerequiredtoassignthecopyrightintheirarticletothePublisher.AcompletedandsignedCopyrightFormmustbelodgedwiththePublisherbeforethearticlecanbepublished.IfyourinstitutionorfundingbodyrequiresthatyoudepositacopyofthepaperinanopenrepositoryoryouwishyourpapertobefreelyavailableweofferanOpenAccesspublishingmodel.Pleasevisitthejournalswebsitewww.impublications.comjnirsformoreinformation.Pleasenotethatyoushouldnotrequestthisuntilafteryourpaperhasbeenacceptedforpublication.SubmissionofarticlesYourarticleshouldbesubmittedviaouronlinesystematwww.impublications.comauthors.IfyoupreferitmaybesenttotheEditororoneoftheRegionalEditorswhoseaddressesappearbelow.PleaseusetheCheckListattheendoftheseinstructionsbeforepostingyourarticle.Receiptofthearticleswillbeacknowledgedandthearticlewillbesubjectedtoapeerreviewprocessshouldanyamend-mentsbenecessaryyouwillbeaskedtomakethesebeforethearticleisacceptedforpublication.Afteracceptanceyourarticlewillbecopy-editedtypesetandproofedtothefirstnamedauthorunlessotherwiserequested.Theproofshouldbecheckedcarefullyfortypo-graphicalerrorsandtoansweranyquestionsraisedbytheCopyEditor.Theproofstageisnotanopportunityfortheauthortomakechangestothearticle.Anysignicantchangesfromtheoriginalacceptedmanuscriptmayonlybeacceptedonpaymentofthecorrectioncosts.AddressesEDITOR-IN-CHIEFProfessorGraemeD.BattenSeaSpecPtyLtdPOBox487WoolgoolgaNSW2456Australia.PhoneFax61-266-562288.E-mailthebattensbigpond.comAuthorsdistantfromAustraliaarewelcomeiftheyprefertosendtheirmanuscripttoIMPublicationsLLP6CharltonMillCharltonChichesterWestSussexPO180HYUKwhowillbepleasedtoforwardittoGraemeBatten.EDITORSDrFranklinE.BartonIILightLightSolutionsLLCPOBox81486AthensGA30608-1486USAE-mailbartonlightlightsolutions.com.DrG.DowneyAshtownFoodResearchCentreAshtownDublin15IrelandE-mailgerard.downeyteagasc.ie.ProfessorT.FearnDepartmentofStatisticalScienceUniversityCollegeLondonGowerStreetLondonWC1E6BTUK.E-mailtomstats.ucl.ac.uk.ProfessorC.W.HuckInstituteofAnalyticalChemistryandRadiochemistryCCB-CenterforChemistryandBiomedicineLeopold-FranzensUniversityInnrain80-826020InnsbruckAustriaE-mailchristian.w.huckuibk.ac.atProfessorS.KawanoKagoshimaUniversityJapan.E-mailkawanoagri.kagoshima-u.ac.jp.DrRogerMederTropicalProductionForestryCSIRO306CarmodyRoadStLuciaQueensland4067Australia.E-mailroger.medercsiro.au.MrK.H.NorrisConsultant11204MontgomeryRoadBeltsvilleMD20705USAE-mailknnirsgmail.com.DrBradSwarbrickCAMOSoftwareASUSAE-mailbswarbrickcamo.com.Checklist1.Arepagesandlinesnumbered2.Isthetextdoublelinespaced3.Areillustrationspreparedaccordingtotheinstructions4.Isthereanabstract5.Aretherekeywords6.Arealltheguresrefereedtointhetext7.Areallthereferencesrefereedtointhetextandinnumericalorderaccordingtotheirrstcitation8.Arethereferencecitationsformedaccordingtotheinstructions9.Arethetablespreparedaccordingtotheinstructions10.Havetheequationsbeencarefullychecked11.Haveyoupreparedtextandillustrationlesinelectronicformat12.Areyoupreparedtopayforthecostofprintingcolourphotographsorillustrations13.Isthereasignedcopyrightform14.Haveyouprovidedfullcontactinformationmailingaddresstelephonefaxnumberande-mailaddress. IMPublicationsLLP6CharltonMillCharltonChichesterWestSussexPO180HYUnitedKingdomT44-1243-811334F44-1243-811711Einfoimpublications.co.ukWwww.impublications.comMembersI.MichaelandK.MichaelPartnershipNo.OC328679VATRegistrationGB917019140CopyrightTransferandPublicationAgreementInordertopublishyourarticleweneedyouragreementinwriting.Pleasetakeamomenttoreadthetermsofthisagreementsigntheformandreturnittousasquicklyaspossible.NameofPublication...................................................................................................................................................................................TitleofArticle...................................................................................................................................................................................NamesofAuthors1...............................................................................Email..............................................................................2...............................................................................Email..............................................................................3...............................................................................Email..............................................................................4...............................................................................Email..............................................................................NameofCopyrightOwnerifnotAuthors...................................................................................................................................................................................Correspondingauthorsaddress............................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................1.BysigningthisformyoutheauthorsorothercopyrightownerherebytransferyourcopyrighttousIMPublicationsLLP.InparticularthismeansthatyougrantustheexclusiverightforthefulltermofcopyrightandanyrenewalsextensionsthereofbothtoreproduceandordistributeyourarticleincludingtheabstractourselvesthroughouttheworldinprintedelectronicoranyothermediumandinturntoauthoriseothersincludingReproductionRightsOrganisationssuchastheCopyrightLicensingAgencyandtheCopyrightClearanceCentretodothesame.Youagreethatwemaypublishyourarticleinthepublicationnamedaboveandthatwemaysellordistributeitwithinthepublicationonitsownorwithotherrelatedmaterial.Inreturnweagreetopublishyourarticleatourexpense.2.Youwarranti.e.promisethatthearticleisyouroriginalworkhasnotpreviouslybeenpublishedandisnotcurrentlyunderconsiderationbyanotherpublication.Ifitcontainsmaterialwhichissomeoneelsescopyrightyoupromisethatyouhaveobtainedtheunrestrictedpermissionofthecopyrightownerpleaseattachevidenceofthisandthatthematerialisclearlyidentifiedandacknowledgedwithinthetext.Youalsopromisethatthearticledoesnottothebestofyourknowledgecontainanythingwhichislibellousillegalorinfringesanyonescopyrightorotherrights.3.Wepromisethatwewillrespectyourrightsastheauthors.Thatiswewillmakesurethatyournamesisarealwaysclearlyassociatedwiththearticleandwhileyoudoallowustomakenecessaryeditorialchangeswewillnotmakeanysubstantialalterationtoyourarticlewithoutconsultingyou.Whenthearticleisacceptedweundertaketopublishitassoonaswereasonablycan.Ifitisrejectedthisagreementisautomaticallycancelledandalltherightscomebacktoyou.Onpublicationwewillsendyouafreecopy.Wewilldoeverythingwereasonablycantomaximisethevisibilityofthepublicationandofyourarticlewithinit. 4.Copyrightistransferredtousandwewillindicatethisinthecopyrightlinewhichappearsonyourarticle.Howeverwegrantbacktoyoutherighttouseyourownarticleprovidedyouacknowledgethepublishedoriginalinstandardbibliographiccitationforminthefollowingwaysaslongasyoudonotsellitorgiveitawayinwayswhichwouldconflictdirectlywithourcommercialbusinessinterests.Youarefreetouseyourarticlefortheinternaleducationalorotherpurposesofyourowninstitutionorcompanymountedonyourownoryourinstitutionsinternalwebsitebutnotonanypubliclyaccessiblewebsiteorothermethodofdistributionorinwholeorinpartasthebasisforyourownfurtherpublicationsorspokenpresentations.SignaturesofAuthors1...............................................................................Date.................................................................................2...............................................................................Date................................................................................3...............................................................................Date.................................................................................4...............................................................................Date.................................................................................NBEverynamedauthormustsignthisformoracopyofitPleasereturnthisformbypostorfaxtoIMPublicationsLLP6CharltonMillCharltonChichesterWestSussexPO180HYUnitedKingdom.Fax441243811711PleasetickifyoudoNOTwantyourpaperpublishedasaJustPublishedPaperbeforecopyeditingandtypesettingiscompleted.Ifthearticleisrejectedthisassignmentisnullandvoid IMPublicationsLLP6CharltonMillCharltonChichesterWestSussexPO180HYUnitedKingdomwww.impublications.com