makingspotflux

VOLUME84NUMBER482DECEMBER2003PAGES529–544Eos,Vol.
84,No.
48,2December2003EOS,TRANSACTIONS,AMERICANGEOPHYSICALUNIONTheAmeriFluxnetworkofeddyfluxcovari-ancetowerswasestablishedtoquantifyvaria-tionincarbondioxideandwatervaporexchangebetweenterrestrialecosystemsandtheatmos-phere,andtounderstandtheunderlyingmech-anismsresponsibleforobservedfluxesandcarbonpools.
ThenetworkisprimarilyfundedbytheU.
S.
DepartmentofEnergy,NASA,theNationalOceanicandAtmosphericAdminis-tration,andtheNationalScienceFoundation.
Similarregionalnetworkselsewhereintheworld—forexample,CarboEurope,AsiaFlux,OzFlux,andFluxnetCanada—participateinsynthesisactivitiesacrosslargergeographicareas[Baldocchietal.
,2001;Lawetal.
,2002].
TheexistingAmeriFluxnetworkwillalsoformabackboneof"Tier4"intensivemeasure-mentsitesasonecomponentofafour-tieredcarbonobservationnetworkwithintheNorthAmericanCarbonProgram(NACP).
TheNACPseekstoprovidelong-term,mechanisticallydetailed,spatiallyresolvedcarbonfluxesacrossNorthAmerica[WofsyandHarriss,2002].
Forbothoftheseroles,theAmeriFluxnetworkshouldbeecologicallyrepresentativeoftheenvironmentscontainedwithinthegeographicboundariesoftheprogram.
Anewecoregion-scaleanalysisoftheexistingAmeriFluxnet-workrevealsthat,whilecentralcontinentalenvironmentsarewell-represented,additionalfluxtowersareneededtorepresentenvironmentalPAGES529,535NewAnalysisRevealsRepresentativenessoftheAmeriFluxNetworkFig.
1.
Therepresentativenessofanexistingspatialarrayofsamplelocationsorstudysites—forexample,theAmeriFluxnetworkofcarbondioxideeddyfluxcovariancetowers—canbemappedrelativetoasetofquantitativeecoregions,suggestinglocationsforadditionalsamplesorsites.
Distanceindataspacetotheclosestecoregioncontainingasitequantifieshowwellanexistingnetworkrepresentseachecoregioninthemap.
Environmentsindarkerecoregionsarepoorlyrepresentedbythisnetwork.
BYWILLIAMW.
HARGROVE,FORRESTM.
HOFFMAN,ANDBEVERLYE.
LAWEos,Vol.
84,No.
48,2December2003combinationsinsouthTexas,theSonoranDesert,andthePacificNorthwest(Figure1).
Analysisoftherepresentativenessofanet-workofsitesentailsacomparisonofdistancesamongcentroidsinamultivariateenvironmental"dataspace.
"Anetworkinthissenseconsistsofageographicconstellationofinstallationsorfacilities,orcansimplyrepresentlocationswheresampleshavebeentaken.
Toquantifynetworkcoverage,wedeterminehowdifferenteachecoregioninthemapisfromtheecore-gioncontainingthemostsimilarnetworksiteorsample.
Thenewanalysisisbasedonasetofstatisti-callyderivedecoregionsproducedforthelower48UnitedStatesusingamultivariateclusteringprocess.
Twenty-fiveenvironmentalconditionsformamultivariatedescriptionoftheenvironmentpresentwithineach1-kmmaprastercell.
Becauseofthesmallcellsizeandthelargenumberofcells,theclusteranalysiswasperformedinparallelonasuper-computer[Hargroveetal.
,2001].
Normalizedvariablevaluesfromeachmapcellareusedascoordinatestoplotthecellsinanenviron-mentalspacewithasmanyaxesastherearemultivariateenvironmentaldimensions.
Twomapcellsthatareplottedclosetooneanotherindataspacewillhavesimilarmixturesofenvironmentalconditions,andarelikelytobeclassifiedinthesamehomogeneousecoregioncluster.
Theusercanspecifythenumberofclusteredecoregionsthatresultfromtheprocess,makingitpossibletodividethemapintoafewlarge,coarselydefinedecoregions,oralargernumberofsmall,highlyspecifiedecoregions(seehttp://geobabble.
ornl.
gov/cgi-bin/pzsforaspectrumofquantitativeU.
S.
ecoregions).
Suchstatisticallygeneratedecoregionscanbeproducedbasedonuser-selectedcontinuousvariables,allowingcustomizedregionstobegeneratedforanyspecificproblem[HargroveandHoffman,2003].
Aniterativek-meansclusteringprocedureassignseachmapcelltotheclosestofkcen-troids.
Attheendofeachiteration,centroidpositionsarerecomputedandanotherassign-mentiterationbegins.
Afterconvergence,themapcells,withtheirnewecoregionassignments,arere-assembledingeographicspace.
Allfinalquantitativeecoregionscontainroughlyequalenvironmentalheterogeneity.
Coordinatesofthefinalcentroidsquantitativelydefinethesynopticconditionsforeachecoregion.
Foreachecoregioninthemap,wefindtheEuclideandistanceindataspacetothesingleclosestecoregionthatcontainsasitefromthenetwork.
Thisdistanceiscodedtoagraylevel,sothatdarkerareasrepresentareasthatarepoorlyrepresentedbytheexistingnetwork(Fig-ure1).
Networkanalysisshowshowwellthesampledenvironmentsrepresenttherestofthemapandidentifiesthebestlocationsfornewsitesorinstallations.
Thebestlocationforanadditionalsitewillbeinplacesthataretheleastwell-represented(darkest)bythenet-workofexistingsites.
The2000mostdifferentmultivariateecore-gionswithintheconterminousUnitedStates(Figure2)werestatisticallydelineatedusing25primaryenvironmentalforcingfactors,includingelevation,meanandextremesofannualtemperature,meanmonthlyprecipita-tion,soilnitrogen,organicmatter,watercapac-ity,frost-freedays,soilbulkdensityanddepth,andsolaraspectandinsolation.
Nodirectveg-etationparameterswereincludedintheseanalyses,onlyprimaryclimaticandphysiographicdrivers.
Thisfinedivisionintomanytightlydefinedecoregionsisfarmorethancanbedelineatedtraditionallyusinghumanexpertise.
Thisfine-scaleecoregionmapwasusedasthebasisforthequantitativeanalysisofnetworks.
Environmentsinthecentral,midwestern,andnortheasternportionsoftheUnitedStatesarewell-representedbyexistingAmeriFluxtowersites(Figure1).
Southern,southwestern,andPacificNorthwestenvironmentsarelesswellrepresentedbyexistingtowersites.
AsaresultofaMay2003proposalsolicitation(DE-FG01-03ER03-22),DOEiscurrentlyselectingadditionaltowersitesinthe"upperMidwestregionoftheUSA,boundedbyMinnesota/Wisconsinonthenorth,Missouri/Oklahomaonthesouth,IndianaontheeastandNebraskaonthewest.
"Thisregionhascomparativelysimpleterrain,alargerangeofseasonalfluxesfromagriculture,mod-estandwidelydistributedanthropogenicemissions,anddensepre-existingmeteorolog-icalmeasurements.
Intermsofrepresentingenvironmentalconditionswithinthelower48statesoftheUnitedStates,however,ourregionalanalysissuggeststhatadditionalsitesinthePacificNorthwestorsouthTexaswouldcon-tributesignificantlymoremarginalrepresenta-tionthanwouldadditionalMidwesternsites.
ShowingGeographicalRepresentationThegeographicrepresentationcontributedbyeachsitecanbeshownbyrepeatingthenetworkanalysiswitheachsitealone.
Impor-tancevaluescanbecalculatedforeachsitebasedonthemarginalrepresentationitaddstothenetwork.
Wereitnecessarytoremoveasite,quantifyingthecontributionofeachsitetonetworkrepresentationcouldminimizetheimpactofsiteeliminationonnetworkrepre-sentation.
Finally,foranetworkthathasnotyetbeendeployed(e.
g.
,NSF'sproposedNationalEcologicalObservationNetwork(NEON)),atheoreticallyoptimalnetworkcanbedesignedthathasthehighestpossiblerepresentationofenvironmentalconditionsonamap,givenaspecifiednumberofsitesandunderlyingecoregions.
Networkanalysisisecoregion-basedandoperatesatthescaleoftheentiresamplingnetwork,consideringhowwellasamplingnetworkrepresentsthemapthatcontainsit.
Itdoesnotconsiderspecificlocalconditions,landuses,disturbancehistory,oranthropogenictreatments—thatis,clear-cuts,forestplantations,oragriculture—unlessdataaboutsuchfine-scalelandconditionsareavailable,andareincludedasinputs.
Networkanalysisdependsonjudiciousselectionoftheenvironmentalvariablesbeingconsidered.
Resultsarecalcu-latedwithrespecttotheselectedinputvari-ablesonlyanddependonthequalityoftheinputdata.
Resultsshowingthenationalrepresentative-nessofNSF'sLong-TermEcologicalResearch(LTER)networkandthenetworkofU.
S.
NationalParkscanbeseenathttp://research.
esd.
ornl.
gov/~hnw/networks/presentation.
Somepoorlyrepresentedecoregionsappearinmanyofthenetworkanalysismaps;thesemaybeoutlyingclusterecoregionsthatarefarfromthecenterofoccupieddataspace.
These"unusual"ecoregionsaregenerallylocatedaroundthegeographicFig.
2.
DividingtheUnitedStatesintothe2000mostdifferentmultivariateecoregionsbasedon25climatic,physiographic,andedaphicfactorsprovidesafoundationfortheanalysisofnetworks.
Theusercanspecifythenumberofclusteredecoregionsthatareproduced.
Allecore-gionswithinthemaphaveroughlyequalenvironmentalheterogeneity.
Eos,Vol.
84,No.
48,2December2003peripheryoftheUnitedStates.
Themore"aver-age"continentalenvironmentsareeasiertosampleandrepresentthanenvironmentsaroundtheedgesofthenation.
ThePacificNorthwest,forexample,isfinelydividedintomanysmallecoregions,makingthiscomplexareahardertocapturewithnetworksites[LawandWaring,1994].
QuantifyingtheSimilarityofEcoregionsLikeanetwork,eachsingleecoregioninthemapbearssomeresemblancetoallotherecoregions.
Similarityofanyecoregiontoallotherecoregionscanbequantifiedanddis-playedasa"representativeness"map.
Onecanselectasingleecoregionofinterest,andthenproduceasortedlistofthesimilarityofallotherecoregionstotheoneselected.
Thechosenecoregionestablishesanoriginindataspace,and,usingtheEuclideandistancefromthisorigintothecentroidofeveryotherecoregion,pair-wisesimilaritymeasurescanbecalculated.
Codingthesepair-wisesimilarityvaluesasgraylevels,thedegreeofsimilarityofallecoregionstotheselectedecoregioncanbemapped.
Mapsshowingtheinnatesimilaritybetweenaparticularecoregionandtherestofthemapcanbeproduced.
Forexample,startingwithaslightlycoarsermapofthe1000mostdifferentecoregionsbasedonthe25primaryenviron-mentalfactorsdescribedearlier,amapof"SmokyMountains-ness"wasproducedthatshowshowsimilarotherregionsaretotheGreatSmokyMountainsofTennesseeandNorthCarolina(Figure3).
ThemultivariateecoregioncontainingtheGreatSmokyMoun-tainsisshowninred.
Darkerareasaremostsimilartotheselectedecoregion.
Themapof"SmokyMountains-ness"statisti-callyrediscoverstheU.
S.
portionofthetem-peratedeciduousforestbiome.
OnesmallspotintheOzarksandonespotintheMonongahelaNationalForestofWestVirginia,thoughspatiallydisjoint,areactuallyapartofthesameecoregionthatcontainstheSmokyMountains,and,atthislevelofdivision,consistofpure"SmokyMountains-ness.
"TheAdiron-dackMountainsofNewYorkarealsorelativelyhighinmultivariate"SmokyMountains-ness.
"Thesecomparativerepresentationmapsquan-tifytheresemblanceamongecoregions.
Representationcanbequantifiedwithinsingleecoregionsaswell.
Therepresentationofindi-vidualmapcells,asindicatedbytheirdistancesfromthecentroidoftheecoregiontowhichtheyareassigned,couldbeusedtopinpointthebestlocationofanadditionalnetworksitewithinthemostpoorlyrepresentedecoregion.
Contoursofquantitativerepresentationwereusedtocharacterizebordersbetweenecore-gionsasgradual,sharp,orofchangingcharacteralongtheirlength[HargroveandHoffman,1999].
WeconsiderthisanalysisoftheAmeriFluxnetworktobeapreliminaryproof-of-concept;manyimprovementscouldbemade.
Acustomecoregionalizationthatisbasedonvariableschosenasspecificallyrelevanttocarbonflux,perhapsincludingdirectvegetationcharacter-istics,shouldbeusedforbestresults.
Expand-ingtheextentoftheanalysistocovertheNorthAmericancontinentortheglobewouldshowhowwelltheexistingAmeriFluxnetworkrepresentsevenlargergeographicareas.
Forexample,inclusionofthesingleAmeriFluxsiteinMexicoandthenumerousCanadianAmeriFluxsitesmightimprovetherepresenta-tionofthenetwork,evenwithinthelower48states.
Representationcouldbeweightedfortheestimatedmagnitudeofthecarbonfluxexpectedfromaregion,orareaweightingcouldbeused,sothatadditionalsitesarenotindicatedasnecessaryinecoregionsthatareunusual,butdonotoccupymucharea.
Usinga"paint-by-the-number"approach,asetofflux-relevantecoregionsmightbeusedasthebasisforextrapolatingmeasurementsmadeatexistingAmeriFluxtowersintoacontinuousgridofseasonalcarbonfluxesacrosstheUnitedStates.
NetworksofinstallationslikeLTER,AmeriFlux,andNEONrepresentsignificantinvestmentsofresearchcapital.
Networkanalysis,asanout-growthofthequantitativetreatmentofecoregions,providesthefirstobjectiveguidanceforthedesign,evaluation,andgrowthofsuchnetworks.
ReferencesBaldocchi,D.
D.
,etal.
,FLUXNET:Anewtooltostudythetemporalandspatialvariabilityofecosystem-scalecarbondioxide,watervaporandenergyfluxdensities,Bull.
Am.
Geol.
Soc.
,82,2415–2434,2001.
Hargrove,W.
W.
,andF.
M.
Hoffman,Usingmultivariateclusteringtocharacterizeecoregionborders,Comp.
Sci.
Eng.
,1,18–25,1999.
Hargrove,W.
W.
,andF.
M.
Hoffman,Quantitativemeth-odstoextendtheecoregionconceptbeyondthelimitsofhumansubjectivity,Environ.
Manage.
,inpress.
Hargrove,W.
W.
,F.
M.
Hoffman,andT.
L.
Sterling,Thedo-it-yourselfsupercomputer,Sci.
Am.
,256,72–79,2001.
Law,B.
E.
,andR.
H.
Waring,Combiningremotesens-ingandclimaticdatatoestimatenetprimarypro-ductionacrossOregon,Ecol.
Appl.
,4,717–728,1994.
Law,B.
E.
,etal.
,Carbondioxideandwatervaporexchangeofterrestrialvegetationinresponsetoenvironment,Agric.
ForestMeteorol.
,113,97–120,2002.
Wofsy,S.
C.
,andR.
C.
Harriss,TheNorthAmericanCarbonProgram(NACP).
ReportoftheNACPCommitteeoftheU.
S.
InteragencyCarbonCycleScienceProgram.
56pp.
,2002.
(Availableathttp://www.
esig.
ucar.
edu/nacp.
)AuthorInformationWilliamW.
HargroveandForrestM.
Hoffman,OakRidgeNationalLaboratory,Tenn;andBeverlyE.
Law,OregonStateUniversity,CorvallisForadditionalinformation,contactWilliamW.
Hargrove,EnvironmentalSciencesDivision,OakRidgeNationalLaboratory,P.
O.
Box2008,M.
S.
6407,OakRidge,TN37831-6407USA;E-mail:hnw@fire.
esd.
ornl.
govFig.
3.
Mapsofsimilaritytoanyselectedquantitativeecoregioncanbeproduced.
TheEuclideandistanceindataspacefromthecentroidofeachstatisticalecoregiontothecentroidofthecho-senregioniscalculated.
Ecoregionswithclosercentroidsaremoresimilarandarecoloreddarkergray.
ThemultivariateecoregioncontainingtheGreatSmokyMountains(showninred)wasselectedfromthis1000-ecoregionmapbasedon25primaryenvironmentalvariables,sothatthemapshowsthequantitativedegreeof"SmokyMountains-ness"acrossthemap.
Thetemperatedeciduousforestbiomeisrevealedtobehighinmultivariate"SmokyMountains-ness.
"