EvaluationanddownscalingofCMIP5climatesimulationsfortheSoutheastUS

FINALPROJECTMEMORANDUM

August21,2015

1.ADMINISTRATIVEProjecttitle:EvaluationanddownscalingofCMIP5climatesimulationsfortheSoutheastUSAgreement#:G13AC00407Awardrecipients:OregonStateUniversity(OSU):PhilipMote,DavidRuppUniversityofIdaho(UI):JohnAbatzoglouTimeperiodcoveredbyreport:6/15/2014through2/28/2015Actualtotalcost:$30,0002.PUBLICSUMMARYThisprojecthasgeneratedaseriesoffreelyavailabledatasetsthatprovideprojectionsofclimatechangeatappropriatespatialscalesthatcandirectlyaddressspecificmanagementquestions.Theseclimatechangeprojectionsaretheresultof“downscaling”outputfromglobalclimatemodels(GCMs)thatformedthebasisofmanyconclusionsintheIntergovernmentalPanelonClimateChange(IPCC)AssessmentReport5(AR5).Thedatasetsincludeprojectionsofclimatevariablesinadditiontodailytemperatureandprecipitationsuchassurfacewinds,humidityandsolarradiationthatareneededinhydrologicandecologicalmodeling.Twoproducts,oneata4-kmresolution,theotherata6-kmresolution,coverthecontinentalUnitedStateshavebeencompletedandareavailablethroughdataserversincludinghttps://www.northwestknowledge.net/Moreover,anevaluationwasdoneofhowwelltheGCMsreproducethehistoricalclimateofSoutheastUSandsurroundingregion.Thisevaluationcanbeusedasonesourceofinformationwhenauserisfacedwithselectingasmallnumberofclimateprojectionsfromthelargersetofavailableprojectionsforanimpactsassessment.Collectively,theguidanceonthecredibilityofGCMsoverthesoutheasternUSandthedownscaleddatasetsprovidenecessaryinformationanddatatodevelopstrategiesforcopingwithclimatechange.3.TECHNICALSUMMARYDownscalingmethodsareusedtobridgethespatialmismatchandbiasesbetweenoutputfromglobalclimatemodels(GCMs–typicalspatialresolutionisseveraldegreeslatitudebylongitude)andinputrequiredbysecondarymodelingapplications.WeadvancedseveraldetailsofstatisticaldownscalingtofacilitatethatdownscaleddatarepresentthesignalofchangesassimulatedbytheGCMwhileretainingmanyofthepropertiesofthetrainingdatasetstoensurecompatibilityforimpactsmodeling.InadditiontodownscalingGCMs,weevaluatedtheGCMswithrespecttotheirabilitytoreproducetheobserved20thcenturyclimatefortheSoutheastUnitedStates(US)andsurroundings.Asuiteofstatisticsthatcharacterizevariousaspectsoftheregionalclimatewascalculatedfrombothmodelsimulationsandobservation-baseddatasets.Lastly,GCMsmodelswererankedbytheirfidelitytotheobservations.

4.PURPOSEANDOBJECTIVES:GCMDownscalingThefirstobjectiveofthisprojectwastogenerateamorephysicallyconsistentanddetailedsetofprojectedmeteorologicalvariablesthanfoundinexistingdownscaledclimateprojections.WhilepreviousdownscalingeffortssuchasBias-CorrectionStatisticalDisaggregation(BCSD)andBias-CorrectionConstructedAnalogs(BCCA)areundoubtedlyvaluable,theyhavelimitationsordrawbacksthatmaymakethemlessdesirableforparticularuses.Theselimitationsincludearestrictedsetofvariables(typicallyonlytemperatureandprecipitation),inabilitytoutilizedailyGCMoutputandpreserveco-variabilityacrossvariables,andissuesinvolvingthetreatmentofmodelbiases.TheMultivariateAdaptiveConstructedAnalogues(MACA,AbatzoglouandBrown2012),andaugmentationstherefore(HegewischandAbatzoglou,forthcoming)largelyovercomesomeoftheselimitationsandallowforamorecomprehensivesetofdownscaledclimateproducts.However,MACAisnotapanaceafordownscaling,asitcannot‘correct’foraglobalclimatemodel’sdeficiencyinsimulatingspatialpatternsofconvectiveprecipitation,orresolvechangesinclimatethatarisefromatmosphere-surfacefeedbacks.Likewise,resolvingthespatialdetailsofconvectivelydrivenprecipitationischallengingforalldownscalingmethods.GCMEvaluationClimatesimulationsfromglobalclimatemodels(GCMs)areoftenreliedupontoprovideplausiblefutureclimatescenariosinclimatechangeimpactsassessmentsatregionalandlocalscales.Frequently,usersareconstrainedtoselectasubsetofthemanyclimateprojectionsavailablefromalargesuiteofGCMs.Modelfidelityisonecriterionthatmaybeusedtoweanthelargepoolofavailableprojections.OursecondobjectiveoftheprojectwastoaidusersinselectingGCMsimulationsbyevaluatinghowindividualGCMsperformedwithrespecttoreproducingthehistorical20th-centuryclimateofthesoutheastUSA.5.ORGANIZATIONANDAPPROACHGCMDownscalingClimatescenariosfrom20CMIP5GCMswiththerequisitedailydatawerestatisticallydownscaledusingtheMultivariateAdaptiveConstructedAnalogues(MACA,AbatzoglouandBrown2012),1950-2005forhistoricalrunsand2006-2099forRCP4.5and8.5(Table1).OutputsfromtwoGCMsthathad360-dayyearswererescaledtoconformtoa365-dayyearcalendar.Twodownscalingproductswereproduced:macav2livnehandmacav2metdata.First,buildingoffthedownscalingperformedfortheNorthwestClimateScienceCenter(whichused“training”datafromthesurfacegriddedmeteorologicaldatasetofLivnehetal.(2013)at1/16th-degreeresolution),weexpandedthedownscalingdomainfromtheNorthwestUStothecontiguousUnitedStatestocreatethemacav2livnehdownscaledproduct.Second,utilizingthe‘training’datafromthegriddedmeteorologicaldatasetofAbatzoglou(2013),thatincludesadditionalvariablessuchasdownwardshortwaveradiationandthesurface,humidity,and10-mwindvelocityatacommon1/24th-degreespatialresolution,wecreatedthemacav2metdataproduct.ThelistofvariablesthatareavailablefromtheseproductsisprovidedinTable2.For thiswork,we augmented the originalMACA downscaling approach to better address some of thebiases inherent in GCMs. The updates included (i) continuous trend preservation of the original GCMsignalusinga31-year,21-daysmoothingwindow,2)useofareducedsetofanalogpatternsbutinclusionof a residual term from the constructed analogs, and 3) joint bias correction of temperature andprecipitation to remove intermodelbiases in temperaturecoincidentwithprecipitation (Hegewischand

Abatzoglou, forthcoming). Thesemodifications resulted in significant improvements in downscaling, asseeninacross-validationstudy.Insummary,MACAwaschosenforthisdownscalingoverothermethodsforthefollowingreasons:•MACAusesdailyoutputfromGCMsandismorereadilyabletocapturechangesinhigher-orderclimatestatistics(e.g.,extremes)thanmethodsthattemporallydisaggregatefrommonthlyprojections.• The spatial downscaling from MACA uses observed spatial patterns rather than using interpolationapproaches.•MACA can be extended tomultiple variables.We downscaled daily temperature, precipitation,windspeed,downwardshortwaveradiationandhumidity.•MACAdownscales someof the variables in sets in order to preserve the dependencies between thevariables. For example, the downscaling of temperature jointly with precipitation has been seen toproducebetterresultsincapturinghistoricalstatisticsofsnowfallandcorrectformodelbiasesspecifictoprecipitatingdaysandthusprecipitationphase.GCMEvaluationRetrospective(i.e.,20thcentury)climatescenariosfrom41CMIP5GCMswereexamined.Wecomparedrelevant20th-centuryobservationswith thesuiteofCMIP5globalmodel resultsaccordingtoasuiteofmetrics designed to determine their suitability for Southeast climate studies following the proceduresoutlinedbyRuppetal.(2013).The metrics listed in Table 1 were calculated from up to 5 observational datasets and all the GCM-simulated datasets of temperature and precipitation. The GCMswere then ranked according to theiroverallfidelitywithrespecttoobservations.6.PROJECTRESULTSGCMDownscalingBetweenthetwodownscalingproductsofmacav2livnehandmacav2metdata,atotalof26terabytesofdownscaleddatawereproduced.Althoughtherearenumerouswaystoanalyzethedata,weprovideacoupleexamplesherethatcanbeexplored infurtherdetail throughourwebpage(seeSec.9).Figure1shows the 20-model mean projected change in Mar-May downward shortwave radiation andprecipitation for years 2070-2099 of experiment RCP 8.5with respect to late 20th century climatology.Figure 2 shows differential rates of warming between the coldest day of thewinter andmeanwintertemperature. This elucidates the additional type of information that can be gleaned from MACAdownscalingthatincorporatesdailyGCMprojections.GCMEvaluationTheprojectgeneratedalargenumberofclimatemetricsperGCMandobservationaldataset.ThesehavebeenpresentedinfiguresthatmaybeusedtocompareamongGCMsortoassesstheabilityoftheCMIP5modelsasawholetofaithfullysimulatetheclimateofthesoutheasternUS.Asanexample,Figure3givesameans of comparing all GCMs and all metrics at once, and thus can be used as an initial means ofidentifyingGCMs that do poorly in a particularmetric, of set ofmetrics, thatmay be of interest for aparticularuse.AdetailedassessmentoftheGCMswithrespecttoeachmetricisprovidedinthetechnicalreport“AnEvaluationofCMIP520thCenturyClimateSimulationsfortheSoutheastUSA”.

7.ANALYSISANDFINDINGSGCMDownscalingDownscaledclimateprojectionshaveallbeenconvertedtoNetCDFformatusingCFmetadatastandardsto ensure compatibility across platforms.We provide both daily and aggregatedmonthlyNetCDF files,acknowledging the different needs of end users. All datasets have been transferred to the NorthwestKnowledge Network (NKN) including the Regional Approaches to Climate Change (REACCH) subserver.NKNprovidesseveralservicestoaidusersinacquiringthehosteddata.First, NKN provides a data catalog to aid users in finding information about the data, as well as tomanuallydownloadsingledatafiles(orsubsets)fromtheinternetindifferentformats(i.e.ascii,NetCDF).Thedatacatalogsforthe2downscaledproductsare:• http://thredds.northwestknowledge.net:8080/thredds/catalog/NWCSC_INTEGRATED_SCENARIOS_ALL_CLIMATE/macav2livneh/catalog.html• http://reacchpna.org/thredds/reacch_climate_CMIP5_macav2_catalog.htmlSecond,NKNprovidesTHREDDSservicestothehosteddatafiles.THREDDSenablesuserstomoreeasilydownloadspatial/temporalsubsetsof thedata, includingtheuseofOPeNDAPtoextractsubsetsof thedatafromwithintheuser’sfavoritesoftwareprogram(R,MATLAB,Python,IDL,etc.).Lastly, though each of the raw NetCDF files represent only 5 or 10-year time spans of data, NKN hasaggregatedalltheyearlyfilesforeachofthescenarios(historical,rcp45,rcp85)intoasinglepointerfile,whichcanbeusedtoaidusersforaccessingallyearsofthedata.ThroughNKN’sservices,usersareabletodownloadspatialsubsetsofthedata,aswellaseachdailyvariableaggregatedtomonthlyaverages.DatastorageandaccessfortheSoutheastdatasetswouldbedecideduponconsultationwithSECSC.GCMEvaluationTherankingofGMCsisnotastraightforwardendeavor.Anyrankingwillvarywiththeparticularmetric,orsetofmetrics,chosen.Also,insomecases,metricswillbephysicallyrelatedtotheextenttheyprovideredundant information. Finally, we may have low large uncertainties about the accuracy of ourestimationofthemetricitself.Givingconsiderationtothelattertwoissues,werankedthemodelsusingamethodology that accounted for information redundancy and favored themetricswe believedweremore reliable. Using the approach, we found that models from the CCSM/CESM1, CNRM, CMCC,HadGEM2,GISS-E2,andMPI-ESMfamiliesrankedhigherthantheothers(Figure4).Thisoverallranking,however, is provided as a suggestion. Individuals can examine the results presented in the technicalreportandusetheseaguidetoamodelselectionsuitedtotheirparticularneedsandobjectives.9.MANAGEMENTAPPLICATIONSANDPRODUCTSInadditiontodownscalingthedatasets,wehavecreatedawebinterfaceforpotentialdatauserstolearnmore about the downscaling methodology and visualize certain aspects of the datasets athttp://climate.northwestknowledge.net/MACA/Thiswebsiteprovidesseveralvisualizationtools,includingtheabilityforuserstoexaminespatialpatternsof change for the variables that have been downscaled across seasons, variable and scenarios. Thesedecision support toolsareofutilityboth fordirectusersof the climatedatasets, aswell as forgeneraldepictionofprojectionsacrosstheregion.Moreover,thetechnicalreport“AnEvaluationofCMIP520th

CenturyClimateSimulationsfortheSoutheastUSA”willbeavailableonwebsiteoncereporthasobtainedOFRcitation.Wearecurrentlyworkingonadataportaltoaidusersindownloadingspatialsubsetsofthedownscaleddailydata,aswellasaggregationsofeachvariabletomonthlyvalues,informatssuchascsv.10.OUTREACHWehavecontinuedtoupdateourwebpagetoprovidevisualizations,guidanceanddata.PresentationsKatherineHegewisch,JohnAbatzoglou,DavidRupp,PhilMote."StatisticallydownscaledclimatedatausingtheMultivariateAdaptiveConstructedAnalogsapproach"5thannualPacificNorthwestClimateScienceConference(PNWCSC),Sept,2014SeattlePublicationsHegewisch,K.C.,Abatzoglou,J.T.,‘AnimprovedMultivariateAdaptiveConstructedAnalogs(MACA)StatisticalDownscalingMethod’,inpreparation.Rupp,D.E.,2015,An Evaluation of CMIP5 20th Century Climate Simulations for the Southeast USA,USGSOpenFileReportXXXXXReferencesRupp,D.E.,J.T.Abatzoglou,K.C.Hegewisch,andP.W.Mote(2013),EvaluationofCMIP520thcenturyclimatesimulationsforthePacificNorthwestUSA,J.Geophys.Res.Atmos.,118,10,884–10,906,doi:10.1002/jgrd.50843.

Table1.AvailabledownscaledCMIP5globalclimatesimulationsusingMACA,1950-2099,RCP4.5andRCP8.5

BCC-CSM1-1

BCC-CSM1-1-M

BNU-ESM

CanESM2

CCSM4

CNRM-CM5

CSIRO-Mk3-6-0

GFDL-ESM2G

GFDL-ESM2M

HadGEM2-CC

HadGEM2-ES

INMCM4

IPSL-CM5A-LR

IPSL-CM5A-MR

IPSL-CM5B-LR

MIROC5

MIROC-ESM

MIROC-ESM-CHEM

MRI-CGCM3

NorESM1-M

Table2.Downscaledvariables

Variable Abbreviation Height

Temperature,maximum tasmax 2m

Temperature,minimum tasmin 2m

Precipitationrate pr Surface

Relativehumidity,maximum rhsmax 2m

Relativehumidity,minimum rhsmin 2m

Specifichumidity huss 2m

Wind,speed was 10m

Wind,eastward vas 10m

Wind,eastward vas 10m

Downwellingsolarradiation rsds Surface

Table 3. Definitions of global climate performance metrics, the confidence in the metrics for modelranking,andobservationaldatasetsusedbymetric.

Metrica Confidencecategory

Description Observationdatasets

Mean-TMean-P

HighestHighest

Mean annual temperature (T) andprecipitation(P),1960-1999

CRU,PRISM,UDelaware,ERA40d,NCEPd

DTR-MMMc Highest Meandiurnaltemperaturerange,1950-1999 CRUe,PRISMe,NCEP

SeasonAmp-TSeasonAmp-P

HighestHigher

Mean amplitude of seasonal cycle as thedifference between warmest and coldestmonth (T), orwettest and driestmonth (P).Monthly precipitation calculated aspercentageofmeanannualtotal,1960-1999.

CRU,PRISM,UDelaware,ERA40d,NCEPd

SpaceCor-MMM-Tb,cSpaceCor-MMM-Pb,c

HighestHigher

Correlation of simulated with observed themeanspatialpattern,1960-1999.

ERA40,NCEPe

SpaceSD-MMM-Tb,cSpaceSD-MMM-Pb,c

HighestHigher

Standard deviation of the mean spatialpattern, 1960-1999. All standarddeviationsarenormalizedby the standarddeviationoftheobservedpattern.

ERA40,NCEPe

TimeVar.1-TTimeVar.8-T

LowerLowest

Variance of temperature calculated atfrequencies (time periods of aggregation)rangingforN=1and8years,1901-1999.

CRU,PRISM,UDelaware

TimeCV.1-PTimeCV.8-P

LowerLowest

Coefficient of variation (CV) of precipitationcalculated at frequencies (time periods ofaggregation) ranging forN = 1 and 8 wateryears,1902-1999.

CRU,PRISM,UDelaware

TimeVar-MMM-Tc Lower Variance of seasonal mean temperature,1901-1999.

CRU,PRISM,UDelaware

TimeCV-MMM-Pc Lower Coefficient of variation of seasonal meanprecipitation,1901-1999.

CRU,PRISM,UDelaware

Trend-TTrend-P

LowerLowest

Linear trend of annual temperature andprecipitation,1901-1999.

CRU,PRISM,UDelaware

ENSO-TENSO-P

LowerLowest

Correlation of winter temperature andprecipitationwithNiño3.4index,1901-1999.

CRU,PRISM,UDelaware

Hurst-THurst-P

LowestLowest

Hurst exponent using monthly differenceanomalies (T) or fractional anomalies (P),1901-1999.

CRU,PRISM,UDelaware

aUnlessotherwisenoted,metricsareaverageoverSoutheastUS.bExpandeddomain:115°W–50°W,15°N–55°N.cMMMistheseasondesignation:DJF,MAM,JJA,andSON.dTemperatureonlyusedinranking,notprecipitation.eNotusedinranking.

Figure1:Projected20modelmeanchangein(top)Mar-Maydownwardradiationandin(top)Dec-Febprecipitation for years 2070-2099 of experiment RCP8.5 versus the historical climate experiment for1950-2005fromdownscaledCMIP5climatemodeloutputs.

Figure2:Projected20modelmeanchange(indegreesC)in(top)coldestminimumtemperature(TMIN)and(bottom)averageminimumtemperature(TMIN)eachwinter(Dec-Feb)foryears2040-2069ofexperimentRCP8.5versusthehistoricalclimateexperimentformodelyears1971-2000.

Figure3:RelativeerroroftheensemblemeanofeachmetricforeachCMIP5GCM.Modelsareorderedfromleast(left)tomost(right)totalrelativeerror,wheretotalrelativeerroristhesumofrelativeerrorsfromallmetrics.

DTR.SONDTR.JJA

DTR.MAMDTR.DJF

SpaceSD.SON−PSpaceSD.JJA−P

SpaceSD.MAM−PSpaceSD.DJF−PSpaceSD.SON−TSpaceSD.JJA−T

SpaceSD.MAM−TSpaceSD.DJF−TSpaceCor.SON−PSpaceCor.JJA−P

SpaceCor.MAM−PSpaceCor.DJF−PSpaceCor.SON−TSpaceCor.JJA−T

SpaceCor.MAM−TSpaceCor.DJF−TTimeCV.SON−PTimeCV.JJA−P

TimeCV.MAM−PTimeCV.DJF−PTimeVar.SON−TTimeVar.JJA−T

TimeVar.MAM−TTimeVar.DJF−TTimeCV.8yr−PTimeVar.8yr−TTimeCV.1yr−PTimeVar.1yr−T

Hurst−PHurst−TENSO−PENSO−TTrend−PTrend−T

SeasonAmp−PSeasonAmp−T

Mean−PMean−T

CESM

1−CA

M5

CNRM

−CM5−2

MPI−ESM

−LR

CNRM

−CM5

CMCC

−CMS

CESM

1−FAST

CHEM

CMCC

−CM

GISS−E2−H

−CC

CCSM

4GISS−E2−R

−CC

CESM

1−BG

CHa

dGEM

2−ES

MIROC5

MPI−ESM

−MR

CSIRO−M

k3−6−0

EC−EAR

THGISS−E2−H

BNU−

ESM

CESM

1−WAC

CMGISS−E2−R

GFD

L−ES

M2G

FIO−ESM

CanESM

2Ha

dGEM

2−AO

GFD

L−CM

3IPSL−C

M5A−M

RHa

dGEM

2−CC

HadC

M3

NorESM

1−M

IPSL−C

M5A−LR

MRI−C

GCM

3bcc−csm1−1

IPSL−C

M5B−LR

FGOALS−g2

GFD

L−ES

M2M

CMCC

−CES

Minmcm

4MIROC−

ESM−C

HEM

MIROC−

ESM

FGOALS−s2

bcc−csm1−1−m

0.0

0.2

0.4

0.6

0.8

1.0

Figure4:41CMIP5GCMsrankedaccordingtonormalizederrorscorefromEOFanalysisofperformancemetrics.Rankingisbasedonthefirst5principalcomponents(filledbluecircles).Theopensymbolsshowthemodels’errorscoresusingthefirst2,4,andall22principalcomponents(PCs).RelativeerroroftheensemblemeanofeachmetricforeachCMIP5GCM.Modelsareorderedfromleast(left)tomost(right)totalrelativeerror,wheretotalrelativeerroristhesumofrelativeerrorsfromallmetrics.