1

AnadjointbasedforecastimpactfromassimilatingMISRwindsintotheGEOS-5dataassimilationandforecastingsystemKevin J. Mueller1, Junjie Liu1, Will McCarty2, and Ron Gelaro2

1 Jet Propulsion Laboratory, California Institute of Technology; 2 Goddard Space

Flight Center (GSFC)

Corresponding author: Junjie Liu Junjie.Liu@jpl.nasa.gov

https://ntrs.nasa.gov/search.jsp?R=20170010247 2020-07-21T12:32:20+00:00Z

2

Abstract1

Thisstudyexaminesthebenefitofassimilatingcloudmotionvector(CMV)wind2

observationsobtainedfromtheMulti-angleImagingSpectroRadiometer(MISR)3

withinaModern-EraRetrospectiveAnalysisforResearchandApplications-24

(MERRA2)configurationoftheGoddardEarthObservingSystem-5(GEOS-5)model5

DataAssimilationSystem(DAS).Availableinnearrealtime(NRT)andwitha6

recorddatingbackto1999,MISRCMVsboastpole-to-polecoverageandgeometric7

heightassignmentthatiscomplementarytothesuiteofAtmosphericMotion8

Vectors(AMVs)includedintheMERRA2standard.Experimentsspanning9

September-October-Novemberof2014andMarch-April-Mayof2015estimated10

relativeMISRCMVimpactonthe24-hourforecasterrorreductionwithanadjoint11

basedforecastsensitivitymethod.MISRCMVweremoreconsistentlybeneficialand12

providedtwiceaslargeameanforecastbenefitwhenlargeruncertaintieswere13

assignedtothelessaccuratecomponentoftheCMVorientedalongtheMISR14

satellitegroundtrack,asopposedtowhenequaluncertaintieswereassignedtothe15

eastwardandnorthwardcomponentsasinpreviousstudies.Assimilatingonlythe16

cross-trackcomponentprovided60%ofthebenefitofbothcomponents.When17

optimallyassimilated,MISRCMVprovedbroadlybeneficialthroughouttheEarth,18

withgreatestbenefitevidentathighlatitudeswherethereisaconfluenceofmore19

frequentCMVcoverageandgapsincoveragefromotherMERRA2wind20

3

observations.Globally,MISRrepresented1.6%ofthetotalforecastbenefit,whereas21

regionallythatpercentagewasaslargeas3.7%.22

4

1 Introduction1

Atmosphericmotionvectors(AMVs),aproxymeasureofwind,are2

indispensabletoregionalandglobalnumericalweatherprediction(NWP)models3

andanalyses.Derivedfromtrackingcloudorwatervaporfeaturesinsatellite4

imagery,AMVsfillsomecriticalconventionalobservationdatagaps(e.g.,theArctic,5

Antarcticandglobaloceans).However,thereremainregionswherewind6

observationsaresparseorunavailable,notablyinthehighlatitudeband(55-65°7

North/South)betweenAMVsobtainedfromregulargeosynchronous(GEO)8

instrumentimageryandthoseobtainedfromconsecutiveorbitsoflowerearthorbit9

(LEO)instruments.AMVsfromcompositeLEO-GEO(e.g.Lazzaraetal.,2014)and10

fromconstellationsofLEOinstruments(e.g.Bordeetal.,2016)increasingly,butnot11

entirely,mitigatethesegaps.AMVsfromLEOarealsolimitedatlowlevels12

(pressures>700hPa)byconcernsabouttheaccuracyoftheradiometricheights13

assignedthere,whichhaveledmultipleNWPcenterstoexcludelowlevelAMVs14

fromoperationalassimilation(Salonenetal.,2015).Cloudmotionvectorwind15

observationsderivedfromtheMulti-angleImagingSpectroRadiometer(MISR)16

instrumentonboardthepolar-orbitingTerracouldhelpmitigatetheabovecoverage17

gaps(Muelleretal.,2013),sincetheirheightsareretrievedbygeometrictechniques18

andtheircoverageisnearlyglobalandconcentratedinthelowertroposphere.19

20

5

MISRmeasuresreflectedsolarradiationinfourbandsfromonboardthesun-21

synchronousTerrasatelliteatninedistinctviewingzenithanglesincludingnadir22

(0°)andfourangles(26°,46°,60°,and70°)distributedalong-trackbothforward23

andaftrelativetoTerra'sflightdirection.Themotionandheightofunderlying24

cloudfeaturesareobtainedfromasingleMISRoverpassbytrackingtheir25

progressionwithin275mresolution380kmswathwidthredbandimageryover26

the3.5minuteintervalbetweentheinitial70°forwardviewandnadir,andthen27

againforthesameintervalbetweennadirandthefinal70°aftview(Horvathand28

Davies,2001a;Muelleretal.,2013).Asidefromthenadirand70°,athirdview29

angleof26°isnecessarilyusedtodifferentiatebetweenparallaxandalong-track30

cloudmotion.Thisapproachyieldsaprecisegeometricheightandcross-trackwind31

component,butarelativelylessprecisealong-trackwindthatissensitivetothe32

accuracyoffeaturetrackingandgeoregistration(Zongetal,2002;Horvathand33

Davies,2001).TheMISRwindheight,cross-track,andalong-trackcomponentshave34

respectiveprecisionof190m,1.1ms-1,and1.8ms-1(Horvath2013).Theretrieval35

algorithmisattunedtostratocumulus,frequentlytrackingthemdespitethepresence36

ofoverlyingcloudswithlessdistincttexture.Asaresult,theoverwhelming37

majority(>95%)ofMISRCMVsamplingisfoundatlowlevels,andsamplingisfar38

betteroveroceanthanland.39

ThegeometricheightsassignedtoMISRCMVsarenotpronetothesignificant40

difficultieswithradiometricheightsassignedtootherAMVs,whicharerecognized41

asakeylimitationtotheirforecastbenefit(Suetal.,2012).AMVheightuncertainty42

accountsfor70%ofvectorwinddifferencesbetweenothertypesofAMVand43

6

rawinsonde(Veldenetal.,2008).Particularlyuncertainarelowlevelradiometric44

heights(pressures>700hPa)assignedtobrokenorsemitransparentcloudsorin45

regionswherethetemperaturelapserateissmall(e.g.,polarregions)orinverted46

(e.g.,themarineboundarylayer).Inthearctic,wheretheseconditionsaretypical,47

lowlevelAMVsaredeemedunreliable(Keyetal.,2003;Santeketal.,2010).In48

comparisonswithAMVsderivedfromtheGeostationaryOperationalEnvironmental49

Satellite(GOES),MISRCMVshavebeenshowntohavelessbiasedheightsinthe80050

hPa–600hPa(~2-4kmheight)range(Muelleretal.,2017),consistentwithsimilar51

findingsofAMVheightbiasrelativetoreanalysis(Salonenetal.,2015).Atthesame52

time,MISRCMVheightsarepronetouncertaintydistinguishingparallaxfromalong-53

trackmotion,leadingtocorrelationoferrorinthesecomponentsandtendencyto54

overestimatetheheightsofupperlevel(>300hPa(~7km))CMVs,thoughthese55

compriselessthan5%oftotalCMVsampling(Muelleretal.,2017).56

Severalstudieshaveprovidedpreliminaryevaluationsoftheforecastbenefit57

MISRwindsmightprovide.Thefirstofthesestudies,Bakeretal.,2014,employed58

anadjointmethodtoquantifythereductionof24-hourforecasterrorsfrom59

assimilatingMISRCMVamongstasuiteofadditionalobservationswiththeNAVy60

GlobalEnvironmentalModel(NAVGEM)4D-VarDataAssimilationSystem.They61

foundthatMISRwindsreduced24-hourglobalforecasterrors,attributingmuchof62

thaterrorreductiontolesseningarelativedearthoflow-levelwindobservations63

assimilatedbytheirmodel.Yamashita(2014)testedassimilationofMISRCMVin64

additiontoroutineobservationswithinthe4D-VARNWPsystemoftheJapan65

MeteorologicalAgency,andfoundincreasedforecastskilloverall.Cressetal.,2014,66

7

assimilatedMISRCMVwithintheoperational3D-VARnumericalweatherprediction67

(NWP)systemoftheGermanWeatherServiceforsummerandwinterof2010,68

findingabenefittotheanomalycorrelationof500hPageopotentialheights69

betweenforecastandanalysis.70

Thesepreviousstudiesdirectlyassimilatedzonalandmeridionalcomponents71

ofMISRCMVretrievals,withnoexplicitmechanismtocapitalizeonthegreater72

accuracyofthecross-trackcomponentsofwindsreportedbyMISR.Inthisstudy,we73

havedecomposedMISRCMVintoalong-trackandcross-trackinordertoassign74

appropriateuncertaintiestoeachcomponentandalsoexploredtheimpactof75

assimilatingonlythemoreaccuratecross-trackcomponent.Complementingearlier76

studies,weevaluatetheforecastimpactofMISRCMVusingtheModern-Era77

RetrospectiveAnalysisforResearchandApplications-2(MERRA2)3D-VAR78

configurationoftheGoddardEarthObservingSystem-5(GEOS-5)modelData79

AssimilationSystem(DAS)(Gelaroetal.,2017).80

Theremainingsectionsareorganizedasfollows:section2describesthe81

datasets,experimentsanddiagnostictoolsusedtoassessMISRwindsbenefittothe82

analysisandforecast.Section3summarizestheresults,includingcomparisonof83

techniquesandparametersforassimilatingMISRwinds.Section4provides84

discussionandconclusions.85

8

2 Dataandmethods86

2.1 ReviewanduseofMISRCMVproducts87

TwodistinctsourcesofMISRCMVsareemployedinthisstudy,themonthly88

aggregatedMISRLevel3CMVproductandtheLevel2NRTCMVproduct.Thetwo89

productsareavailableonlineandrespectivelytaggedasMI3MCMVNand90

MI2TC_CMV_HDF_NRTattheNASALangleyDistributedActiveArchiveCenter.The91

formerisavailablewith24hourlatencyarchivedbackto2000,thelatterwithNRT92

latency(95%ofCMVSinunder2.5hours)archived30daysbackfrompresent.93

ExcludedfromthisstudyaretwootherMISRproductscontainingCMVlesssuitable94

forassimilation.TheMISRLevel2Cloudproduct(taggedMIL2TCSP)containsa95

supersetofCMVfromtheLevel3CMVproductthatincludesretrievalsoflower96

quality.TheMISRLevel2Stereoproduct(taggedMIL2TCST)containslessaccurate97

CMVretrievedbyalegacyalgorithm.98

99

TheLevel2NRTCMVproductusesthesameretrievalandqualitycontrol100

algorithmsasthestandardLevel3CMVproduct,generatingresultscomparableto101

thelatter.However,thetwoproductshaveminordifferencesowingtodifferences102

betweentheNRTandstandardprocessing(STD)versionsofupstreamLevel0(L0)103

andLevel1(L1)datainputs.TheNRTsoftwarepipelineisappliedtoincomingL0104

instrumentandsatellitedatainsessionsassociatedwithaslittleasfiveminutesof105

data,whereasthestandardpipelineoperatesonthesamedataconsolidatedinto106

sessionscomprisingonefullorbit(90minutes).Withoutthisconsolidation,NRT107

9

processingissubjecttolostcoverageassociatedwithgapsintheavailabilityof108

necessaryinputsattimeofprocessing.Additionally,thecontinuouslyupdated109

recordofcorrectionstocamerapointingusedtoperformin-flightgeometric110

calibrationissensitivetothedifferencesbetweenNRTandSTDprocessing.111

Calibrationdiscrepanciesareresponsibleforaroot-mean-square-vector-difference112

(RMSVD)of3ms-1betweencollocatedLevel2NRTCMVsandLevel3CMVs113

(Muelleretal.,2013b).114

115

2.2 GEOS-5model,assimilationsystem,andadjointmethodology116

Thisstudyemploysversion5.13.0oftheGEOS-5DAS,withrevisionstosupport117

MISRwindsassimilationanddeterminationofadjointsensitivity.Version5.13.0is118

associatedwithofficiallyreleasedGEOS-5dataproductsbetween20September119

2014and1May2015.TheC180(1/2degreelatitude-longitude)gridwith72layers120

betweenthesurfaceand0.01 hPaisemployed,withdefaultversion5.13.0model121

parametersandassimilatedobservations(otherthanMISRCMV)definedbythe122

MERRA-2reanalysis(Gelaroetal.,2017).MERRA-2incorporatesabroadrangeof123

observations,includinggeostationaryAMVsfromGOES,124

ThemeteorologicalanalysisinGEOS-5usestheGridpointStatistical125

Interpolation(GSI)3D-Var[Wuetal.,2002;Purseretal.,2003a,2003b]assimilation126

methods.Theobjectiveoftheassimilationistoproduceananalysisfieldforwhicha127

costfunctionconstructedfromtheobservation-minus-analysis(O-A)residualsis128

minimizedsubjecttoassumedforecastandobservationerrorstatistics[Cohn,129

1997].TheGSIperformsminimizationrelativetocontrolvariablesincludingstream130

10

functioncontributionfromwind,unbalancedvelocitypotentialfunction,unbalanced131

temperature,unbalancedsurfacepressure,moisture,cloudwater,ozone,and132

coefficientsforthebiascorrectionofthesatelliteradiancedata.133

Anadjointsensitivitymethodisemployedtocalculatetheimpactofeach134

individualobservationontheshort-rangeforecastsimultaneously,producing135

resultsthatcanbeeasilyaggregatedbydatatype,location,channel,etc[Gelaroet136

al.,2007;ZhuandGelaro,2008;Gelaroetal.,2010].Itisthesamemethodused137

operationallytomonitorandevaluatetheimpactofassimilatedobservations.138

Impactsaremeasuredrelativeto24-hourand30-hourforecasterrordifferencesin139

totalmoistenergy.Theforecasterrorismeasuredagainstanalysisstateatthe140

verificationtime,t=24hours(LanglandandBaker,2004).The24-hourand30-hour141

forecastareinitiatedattime,t=0hours,andattime,t=-6hours,respectively.The142

differencebetweentheformerandlatterforecastsareduetoobservations143

assimilatedattheanalysistimet=0hours.Themethodisundertakenforevery6-144

hourtimestep,facilitatingimpactassessmentofallassimilatedobservationsineach145

experiment.146

147

2.3 Experiments148

2.3.1 Thinningandscreeningmethods149

MISRCMVarereportedwith17.6×17.6griddedresolutionwithvertical150

coordinatesofheightrelativetotheEarth'sellipsoid.Forthisstudy,thesetofCMV151

reportedpertimestepwasthinnedsuchthatonlyoneCMVcouldbeassimilatedper152

100kmx100kmx100hPavolumeonamodelalignedgrid.Thisthinningincluded153

11

simpletransformationofMISRreportedgeometricheight,h,intopressure154

coordinatesassumingaconstantstandardatmosphere,thatis:155

𝑝 = 𝑝!𝑒!!!

wherek=1.186×10-4m-1,andps=1013.25hPa.Notethatthisheight-pressure156

conversionformulawasonlyusedinthethinningprocess,whilethemodel157

geometricheightandobservationheightwasusedinverticalinterpolationduring158

dataassimilationprocess.Inadditiontospatialthinning,MISRCMVwithheights159

reportedbelowmodelsurfaceelevationorabove15kmwerealsoexcluded.Forall160

theexperimentslistedinsection2.3.3,agross-errorthresholdwasalsoappliedto161

screenCMVdifferingfrombackgroundstatebymorethan8.0ms-1.162

QualityIndicator(QI)valuesassignedtoMISRCMVweregivennoinfluenceon163

theuncertaintyassignedduringassimilation.CMVQIvaluesareameasureof164

retrievalconsistencybetweenneighborsandbetweenredundantforwardandaft165

camerabasedestimates.GermanWeatherServiceexperimentsfoundCMVQIto166

poorlypredictCMVinfluenceonforecastskill(Cress,2014).Consistentwiththeir167

finding,ourownexperimentsshownegligiblecorrelationbetweenperobservation168

MISRCMVforecastimpactandQI.169

170

2.3.2 Assimilationmethods,uncertaintyassignment,andassimilationexperiments171

Table1listsmodelexperiments,threeofwhichdifferinthewaythewindswere172

assimilatedandtheuncertaintywasassignedtoeachMISRCMV,allsharingthe173

sameSeptember-October-November(SON)timeperiodin2014.Inthefirst174

12

experiment,labeledUV,,weassimilateMISRzonal(u)andmeridionalwind(v)175

directly,thesameastheassimilationmethodsemployedinpreviousMISRwind176

assimilationstudies(e.g.,Bakeretal.,2014).Theuncertaintiesappliedwiththis177

methodare3.0ms-1fortheuandvvectorcomponents.Thecomponentsarenot178

treatedindependently,soeitherbothcomponentsarerejectedorneitherduring179

assimilation.Inthesecondexperiment,labeledATCT,eachMISRCMVistranslated180

intoalong-trackandcross-trackcomponentsbasedonviewinggeometrythatare181

thenassimilatedindependentlywithrespectiveuncertaintiesof8.0ms-1and2.0ms-182

1.Thisapproachtakesadvantageofthecross-trackhavinggreatertheoretical183

accuracy(Zongetal,2002;HorvathandDavies,2001).Theassigneduncertaintyfor184

thecross-trackwindisthesameasassessedinvalidationstudies(Horvath2013,185

Muelleretal.,2017).Theuncertaintyof8.0m/sforthealong-trackMISRCMVis186

conservative,beingsetmuchgreaterthantheglobalalong-trackRMSdifference187

relativetoGOESAMV(3.2ms-1)inordertolimitthepotentialinfluenceofalong-188

trackbiasandregionsofgreateruncertainty(Muelleretal.,2017).Inthethird189

experiment,labeledCT,onlythemoreaccuratecross-trackwindcomponentis190

assimilated.Lastly,acontrolexperiment,labeledCONTROL,wasconductedforthe191

sametimeperiod,butwithnoassimilationofMISRCMV.192

Inadditiontotheaboveexperimentscomparingassimilationmethodology,two193

variationsoftheATCTexperimentwereconductedusingthesameassimilation194

approach.Thefirst,labeledNRT,usestheMISRNRTCMVproductratherthanthe195

MISRL3CMVproduct,tocomparetheirrelativeeffectiveness.Thesecond,labeled196

ATCT15extendsthetimespanofouranalysistoMarch-April-May(MAM)of2015.197

13

3 ResultsandDiscussion198

3.1 Overview199

Inthefollowingsections,weinvestigatetheperformanceofthreeassimilation200

approachesthroughcomparisonsamongtheCT,UV,andATCTexperiments(3.2);201

comparethesamplingandnetimpactofMISRCMVrelativetootherassimilated202

observationsfortheATCTandATCT_15experiments(3.3);andassessthe203

consistencyofMISRCMVsamplingandforecastimpactbetweentheNRTCMVand204

thestandardprocessing(3.4).205

3.2 SensitivityofforecastimpacttothreemethodsofassimilatingMISRCMVs206

ComparisonsofmeanforecastbenefitamongATCT,UV,andCTexperiments207

applyingdistinctassimilationmethodologiestothesamedatainputsdemonstrate208

thesuperiorityoftheATCTapproach.AsshowninTable2,themeanimpactper209

six-hourtimestepcontributedbytheassimilatedMISRCMVwas-25±18 mJ kg-1 for210

ATCT,roughlytwicethatofUVand70%greaterthanthatoftheCTexperiment.211

Figures1d,e,andfpresentobservation-minus-forecast(OMF)(six-hour212

forecast)statistics,showcasingnegligibleOMFbiasforassimilatedcross-track213

componentsofMISRCMVduringtheATCTexperiment,butbiasaslargeas2.0ms-1214

forthealong-trackcomponent.Intheabsenceofmodelorobservationbias,OMF215

shouldbeunbiased(e.g.,Kalnay2003).Here,along-trackOMFbiasfollowsa216

patternbroadlyconsistentwiththatseenincomparisonbetweenMISRCMVsand217

GOESAMVs(Muelleretal.,2017).Thealong-trackcomponentbiasatlowerlevels218

(below750hPa)isproportionaltoheight.Thebiaschangesfromnegativeto219

14

positivearoundthepeakheightofsamplingdensity,at850hPainthetoprowof220

Figure1.Thisisaconsequenceofthecorrelation(Zongetal.,2002)betweenerror221

intheheightanderrorinthealong-trackcomponentofMISRCMVscausing222

preferentialsamplingofnegative/positivealong-trackbiasatheightsbelow/above223

thepeakinsampling.Above750hPa,MISRCMValong-trackbiashasnosuch224

gradient,insteadbeingconsistentlyontheorderof1-2ms-1athighlatitudeswhere225

themajorityofsamplingisfound.Overthetropics,wheresamplingissparse,the226

biasisnotevident.Figures1d,1e,and1falsoshowbiasprofilesfortheUV227

experiment,whereinthesameunderlyingpositivealong-trackbiasmanifestsas228

principallysouthwardcomponentbiasatlowlatitudesandbothwestwardand229

southwardbiasatmidtohighlatitudes.230

EvidentinFigures1g,h,andj,heightsandregionswherethealong-trackbiasis231

mostpronouncedintheATCTexperimentalsoshowthelargestdiscrepancies232

betweenUVandATCTintheforecasterrorreduction.Thisismostevidentinthe233

southernextra-tropicswhereassimilationofthevcomponentintheUVexperiment234

actuallydegradestheforecastformid-levelCMVsatheightsfrom900hPato500235

hPa.Incontrast,thealong-trackcomponentintheATCTexperimentisconsistently236

beneficial.Incontrasttothevcomponent,theassimilateducomponentconsistently237

improvestheforecast,buttoafarlesserdegreethantheassimilatedcross-track238

componentintheATCTexperiment,whichislargelyduetothelargeucomponent239

bias(Figure1d,e,andf)andrandomerrorsrelativetocross-trackcomponent(2.5240

m/svs.2.1m/s).ItisalsoevidentinFigures1a,1b,and1cthatthesameheights241

andregionswherethev-componentdegradedtheUVforecastalsoproduceda242

15

greaternumberofrejectedCMVinUVrelativetoATCT.Thenorthernextra-tropics243

exhibitthesametrendsasabove,thoughtoalesserextent.Thetrendsobservedin244

thetropicsrepresentaninstructivecounterpoint.Becausethemeanalong-track245

biasislessconsistentandnotaslargethere,andbecausetheprojectionofcross-246

trackandalong-trackwindsismorealignedwithuandvcomponentsthere,the247

forecastbenefitisroughlyequivalentfortheATCTandUVexperiments.248

FortheCTexperiment,thealong-trackcomponentwasexcludedalltogether.249

Asevidentforalllatitudebands(Figures1g,h,andi),assimilatingCTonly250

consistentlyreducedtheforecasterror,butchoosingnottoassimilatethealong-251

trackcomponenthadanadverseeffectonthebenefitprovidedbythecross-track252

component,especiallyovertheextra-tropics,reducingitsbenefitbyasmuchas253

40%.Riishojgardetal.(2008)alsoshowedthatsinglewindcomponent254

assimilationproducedapoorerrepresentationofthatfieldintheanalysisstate.255

Theyarguedthattheanalysisaccuracyofsinglewindassimilationdependsmoreon256

theforecasterrorcovariancespecifiedindataassimilationcomparedtoassimilating257

twowindcomponents.Stoffelenetal.(2005)andMarseilleetal.(2008)alsoargued258

thatproperbackgrounderrorcovarianceisessentialtomaximizesinglewind259

observationimpact.InCT,weusedthesameforecasterrorcovarianceasinthe260

ATCTcase.ThepoorerperformanceofCTindicatesthatthedefaultforecasterror261

covariancedoesnotaccuratelycapturetherelationshipsbetweendifferent262

dynamicalvariables.Sincethewindandmassfields,suchaspressure,aremore263

tightlycoupledovertheextra-tropicsintheforecasterrorcovariance,the264

16

degradationofassimilatingasinglecomponentislargerovertheextra-tropics265

comparedtothetropics(Figure1g,h,andi).266

3.3 MISRforecastimpactrelativetootherinstruments267

Theglobal24-hourforecastimpactamongstMISRCMVsandotherclassesof268

satelliteinstrumentsiscomparedfortheATCT_15experimentinFigure2a.(The269

resultsofthecomparableATCTexperimentareroughlyequivalent,withminor270

differencesdiscussedinsection3.4).MISRCMVsrepresent1.6%ofthetotal271

forecastbenefitfromalloftheobservations,whileAMVsfromlowerearthorbit272

(LEO)andgeosynchronous(GEO)represent~15%.Thetotalimpactofsatellite273

windsissecondbehindthatofinfrared(IR)andmicrowave(MW)radiance274

observations,whichrepresentjustover50%.Rawinsondeanddropsondeprofiles275

(labeledRAOB+SND)andin-situaircraftmeasurementsthatincludewind,276

temperature,andpressureeachrepresentanother~12%.Surfacewindforcingas277

measuredbymicrowavescatterometersrepresent1%ofimpact.Anadditional278

10%offorecastimpactnotplottedinFigure2aiscontributedbyvariousland-and279

ship-basedobservations,byradiooccultationobservations,andbypilot-balloon280

measurementsofwind.Figure2bshowstheper-obsimpactofMISRCMVtobethe281

largestoftheaboveobservationgroups,withmagnitudecomparabletoRAOB+SND.282

MISRCMVsarebroadlybeneficialeverywhere,withgreatestbenefitevidentat283

highlatitudeswherethereisaconfluenceofmorefrequentCMVcoverageandgaps284

incoveragefromotherwindobservations.Forexample,Figure2cshowsthatMISR285

CMVcontributes3.7%ofthetotalforecastbenefitbetweenlatitudes55°and70°286

South,morethandoublethatoftheglobalmeanforecastimpactof1.6%.Figure3287

17

showscoverageforclassesofwindobservations,wherecoverageisdefinedbythe288

fractionof6hourperiodsduringwhichoneormoreobservationwasassimilated289

withineach2.5°latitude×2.5°longitudegridcell.Figure3ashowsthefrequencyof290

CMVcoverageinthemonthsofSONtobeunder10%atlowlatitudes,andupto291

20%or50%dependingonseasonathighlatitudes.MISRCMVcoverageis292

primarilygovernedbythesatelliterepeatinterval,varyingfrom9daysatthe293

equatortoaslittleas90minutesnearthepoles,andbyseasonalvariationof294

availablesunlightathighlatitudes.GreaterMISRCMVsamplingathighlatitudes295

synergisticallycoincideswithgapsbetweenLEOandGEOAMVcoverageevidentin296

Figure3b,ultimatelyproducinggreaterforecastbenefitforthoseregionsasevident297

inFigure4a.OvertheSouthernOcean,theseregionsalsocoincidewithapaucityof298

aircraftandsondeobservations(Figures3d&3e),and,correspondingly,even299

greaterforecastbenefit.AnotherregionofenhancedMISRCMVbenefit(Figure4a)300

isfoundovercentralAsiainthegapbetweenGEOAMVscapturedfromMeteosat-9301

andMTSAT-2instruments(Figure3b).Theregionalsolacksfrequentcoverage302

fromaircraftandsondes(Figures3e&3f),totheextentthattherareinstancesof303

MISRCMVretrievedtherehaveoutsizedinfluence.Overocean,thegeographic304

distributionofforecastbenefitfromMISRCMVsisrathersimilartothatof305

scatterometerwinds,possiblyreflectingthefactthatbothlargelyorentirely306

providelowlevelconstraintsonthewindfield(Figures3cand4c).Forexample,307

bothprovidenegligiblebenefitoverlargeswathsofthetropicalPacificandAtlantic308

oceans,whichisprimarilyduetothedensewindobservationsfromthedefaultLEO309

andGEOsatellites(Figure4b).AlthoughtheforecastbenefitfromMISRCMVis310

18

evidentlyenhancedwherecoveragefromotherwindobservationsissparse,MISR311

CMVsalsoexhibitsignificantbenefitsinwell-sampledregionssuchastheNorth312

Pacific.313

3.4 SensitivityofMISRwindforecastimpacttoassimilationtimeperiodand314

MISRCMVproducts315

ATCT_15andATCTwerecarriedoutovertwodifferentseasons:borealsummer316

andfallrespectively.Comparingthesetwoexperimentshelpsidentifythesensitivity317

ofMISRwindimpacttotimeofyear,whileaggregatingthemprovidessixmonthsof318

simulation.Onadailybasis,MISRCMVsprovideaconsistent24-hourforecast319

benefitthroughoutATCTandATCT_15asindicatedinFigure5byatimeseriesof320

perorbitforecastimpactwhereintherunningmeanover15orbits(i.e.therough321

equivalentof24hours)isalwaysbeneficial(i.e.negativecontributiontoforecast322

errornorm).Measurementsofforecastimpactareinherentlynoisy,withthe323

standarddeviationofperorbitCMVimpacthavingcomparablemagnitudetothe324

mean,thatis10Jkg-1×10-3.Still,theoverwhelmingmajority(88%)oforbitswith325

MISRCMVsamplingarefoundtoprovideanetforecastbenefit,whiletheinfrequent326

remainderisbroadlydistributed,suchthatnodurationofsequentialorbits327

contributesasignificantregression.Thelargestsingleorbitforecastregressionis328

28.9Jkg-1×10-3duringATCT.AsvisualizedinFigure5andTable2,thenumberof329

MISRCMVassimilatedonaper-orbitbasis(2500)varieslittle(±360)inATCTor330

ATCT_15.NordoesthenumberofCMVrejected(330±110)duringcomputationof331

modelanalysisstatethroughincrementalminimizationoftheGSIcostfunction.332

19

333

ThenearequivalenceofCMVsix-hourlyforecastimpactsinATCTandATCT_15334

isacoincidentalbyproductofATCTproducing16%greaterimpactperorbitoffset335

by25%fewerorbitsproducingvalidCMVsampling.Thegreaterper-orbitCMV336

benefitinATCTcanbetracedtoseasonalityofsampling,withATCThavingagreater337

fractionoftotalsamplinglocatedovertheSouthernOceanwherethemodelmost338

benefitsfromassimilatingCMVs.ThesamplingdeficitinATCTcanbetracedtotwo339

gapsevidentinFigure5thatwerecausedbyatemporarysuboptimalsoftware340

configurationaffectingTerraattitudedatathathadbeenusedintheMISRstandard341

processingchainduringSeptember2014.Theunderlyingissue,whichwas342

identifiedandrectifiedinNovember2014,didnotaffectotherTerrainstrumentsor343

MISRscienceproducts,andalsodidnotaffectMISRNRTprocessing-hencethe344

absenceofgapsinFigure6correspondingtothoseevidentinFigure5.345

346

Figure6showstimeseriesoftheMISRCMVsimpactandobservationcountin347

theNRTexperiment,whichassimilatesMISRNRTCMVs.Relativetostandard348

processing,theNRTCMVsarepronetolossesofsamplingduetotimelinessof349

necessarydatainput.Asaresult,NRTassimilateslesssamplesperorbit(1900)350

withgreatervariabilityofperorbitsampling(±600).RelativetoATCT,thefraction351

ofsamplinginNRT(76%)isconsistentwiththefractionofforecastbenefit(72%)-352

thatis6.8outof9.5Jkg-1×10-3.Thedistributionandmagnitudeofforecast353

regressionsonaperorbitbasisarealsocomparabletoATCT.354

355

20

4 Conclusions356

Aseriesofexperimentshavebeenconducted,demonstratingthebenefitof357

assimilatingcloudmotionvectorsfromtheMISRCMVsoverperiodscovering358

September-October-Novemberof2014andMarch-April-MaywithintheGEOS-5359

DASasdeterminedbyanadjointbasedforecastsensitivitymethod.Whereas360

previousstudieshavedirectlyassimilatedthezonalandmeridionalcomponentsof361

MISRCMVs,thisstudydemonstratesmoreconsistentlybeneficialandtwiceaslarge362

ameanforecastbenefitwhenassimilatingalong-trackandcross-trackcomponents363

andassigninglargeruncertaintiestolessaccuratealong-trackcomponent.Although364

themorecertaincross-trackcomponentcontributesmorethan90%ofthetotal365

forecastbenefitwhenassimilatingbothalong-trackandcross-track,assimilating366

onlythelatterprovidesonly60%oftheforecastbenefitasboth.Systematicalong-367

trackbiasinMISRCMVsconsistentwithearlierstudieswasevidentinOMF368

statistics.Thisfactoredintothebenefitsofassigninggreateruncertaintytothe369

along-track.Anotherapproachworthinvestigatingwouldbeapplicationofa370

height-andpossiblylatitude-dependentalong-trackbiascorrection.371

TheoverallbenefitofoptimallyassimilatingMISRCMVswasa1.6%372

contributiontotheglobalreductionofthemoistenergyerrornormfor24-hour373

forecasts,withabouttwicethatpercentageofcontributioninregionssuchasthe374

SouthernOceanthatarelesswellobserved.Notethattheimpacton24-hour375

forecasterrorreductionisonlyonemeasureoftheobservationimpact.Theoverall376

reductionon24-hourforecasterrorsfromassimilatingMISRwindscorroborates377

21

earlierstudiesshowinganoverallbenefitfromMISRCMVsasmeasuredbyvarious378

metricswithinmultiplemodels(e.g.,Yamashita(2014)).Themagnitudeofbenefit379

ispromisinginregardtothemulti-angleretrievalofCMVs,giventhelimitationson380

MISRcoverageimposedbyitsrelativelynarrow360kmswath.Asinglewider-381

swathmulti-angleimager,atandemconvoyofsuchimagers(whichavoidsthe382

ambiguitybetweenparallaxandalong-trackmotion),oramultitudeoflowcost383

nano-satellitevariantscouldallprovidesignificantlygreaterforecastbenefit.384

5 ACKNOWLEDGMENTS385

WeacknowledgethefundingsupportfromNASAdataforoperationalanalysis386

(NDOA)programundergrantNNH13ZDA001N.Weacknowledgethetechnical387

supportfromJoeStassi(GMAO),DanHoldaway(GMAO),andMetaSienkiewicz388

(GMAO).WeappreciatetheinstructivecommentsanddiscussionswithDr.Nancy389

Baker(NRL),andthesupportfromDr.DaveDiner(JPL),VeljkoJovanovic(JPL),and390

MichaelGaray(JPL).Allthecalculationswerecarriedoutwiththesupercomputer391

fromNASAcenterforclimatesimulations(NCCS).Thisresearchwascarriedoutat392

JetPropulsionLaboratory,CaliforniaInstituteofTechnology,undercontractwith393

theNationalAeronauticsandSpaceAdministration. 394

22

6 REFERENCES1

Baker,N.L.,P.M.Pauley,R.H.Langland,K.Mueller,andD.Wu,2014:Anassessment2

oftheimpactoftheassimilationofNASATERRAMISRatmosphericmotion3

vectorsontheNRLglobalatmosphericpredictionsystem.SecondSymposiumon4

theJointCenterforSatelliteDataAssimilation,94thAmericanMeteorological5

SocietyAnnualMeeting.Atlanta,Georgia,GeorgiaWorldCongressCenter.6

http://www.jcsda.noaa.gov/documents/meetings/AMS2014/2nd_JCSDA_Baker_7

MISR.pptx8

Borde,R.,O.Hautecoeur,andM.Carranza.(2016)EUMETSATGlobalAVHRRWind9

Product.J.Atmos.Ocean.Tech.,133,429-43810

Cohn,S.E.,1997:Anintroductiontoestimationtheory.J.Meteor.Soc.Japan,75,11

257-28812

Cress,A.,(2014)ImprovingtheUseofSatelliteWindsattheGermanWeather13

Service,(2014)TwelfthInternationalWindsWorkshop,Copenhagen,Denmark.14

http://www.eumetsat.int/website/wcm/idc/idcplg?IdcService=GET_FILE&dDo15

cName=PDF_CONF_P61_S4_02_CRESS_V&RevisionSelectionMethod=LatestRele16

ased&Rendition=Web17

Gelaro, R., Y. Zhu, and R. M. Errico, 2007: Examination of various-order adjoint-based 18

approximations of observation impact. Meteorologische Zeitschrift 16, 685–692. 19

Gelaro R, R. H. Langland, S., Pellerin, and R. Todling, 2010: The THORPEX 20

observation impact inter-comparison experiment. Mon. Weather Rev. 138: 4009–21

4025. 22

23

Gelaro, R., and Coauthors, 2017: The Modern-Era Retrospective Analysis for Research 23

and Applications, version 2 (MERRA-2). J. Climate, 30, 5419–5454 24

Horvath, Á. and R. Davies, 2001: Feasibility and error analysis of cloud motion wind 25

extraction from near-simultaneous multiangle MISR measurements, Journal of 26

Atmospheric and Oceanic Technology, 18-4, 591-608 27

Horváth, Á. 2013: Improvements to MISR stereo motion vectors., J. Geophys. Res., 118 28

Kalnay, E. (2003), Atmospheric Modeling, Data Assimilation, and Predictability, 341 29

pp., Cambridge Univ. Press, New York 30

Key, J. R., D. Santek, C. S. Velden, N. Bormann, J.-N. The ́paut, L. P. Riishøjgaard, Y. 31

Zhu, and W. P. Menzel, 2003: Cloud-drift and water vapor winds in the polar regions 32

from MODIS. IEEE Trans. Geosci. Remote Sens., 41, 482–492. 33

Langland, R. H., and N. Baker, 2004: Estimation of observation impact using the NRL 34

atmospheric variational data assimila- tion adjoint system. Tellus, 56, 189–201. 35

Lazzara, M.A., R. ATCTorak, D.A. Santek, B.T. Hoover, C.S. Velden, and J.R. Key, 36

2014. High-Latitude Atmospheric Motion Vectors from Composite Satellite Data. J. 37

Appl. Meteor. Clim. 53, 534-547. 38

Marseille,G.J.,A.Stoffelen,J.Barkmeijer,2008.SensitivityObservingSystem39

Experiment(SOSE)-ANewEffectiveNWP-basedToolinDesigningtheGlobal40

ObservingSystem.Tellus,A,1,60,216-233,10.1111/j.1600-0870.2007.00288.x41

Mueller, K.J., M.J. Garay, C. Moroney, and V. Jovanovic, 2012: “Enhanced MISR Cloud 42

Motion Vectors: Performance Evaluation Relative to RAOB, MODIS, and GOES,” 43

11th International Winds Workshop, Auckland, New Zealand 44

24

Mueller, K.J., C.M. Moroney, V. Jovanovic, M.J. Garay, J-P Muller, L. Di Girolamo, and 45

R. Davies, 2013a: MISR Level 2 Cloud Product Algorithm Theoretical Basis, JPL D-46

73327 47

http://eospso.gsfc.nasa.gov/sites/default/files/atbd/MISR_L2_CLOUD_ATBD-1.pdf 48

Mueller, K.J., C.M. Moroney, and V. Jovanovic, 2013b: "MISR Level 2 Cloud Product 49

Quality Statement, September 14, 2012", 50

https://eosweb.larc.nasa.gov/sites/default/files/project/misr/quality_summaries/L2TC51

_Cloud_Product.pdf 52

Mueller, K.J., D.L. Wu, A. Horvath, V.M. Jovanovic, J.-P. Muller, L.D. DiGirolamo, M.J53

. Garay, D.J. Diner, C.M. Moroney, and S.Wanzong, 2017: Assessment of MISR Clo54

ud Motion Vectors (CMVs) relative to GOES and MODIS Atmospheric Motion Vect55

ors (AMVs). J. Appl. Meteor. Climat. doi:10.1175/JAMC-D-16-0112.1, in press. 56

Purser, R. J., W. Wu, D. F. Parrish, and N. M. Roberts, 2003a: Numerical aspects of the 57

application of recursive filters to variational statistical analysis. Part I: Spatially 58

homogeneous and isotropic Gaussian covariances. Mon. Wea. Rev., 131, 1524–1535. 59

Purser, R.J., W.-S. Wu, D.F. Parrish, and N.M. Roberts, 2003b: Numerical aspects of the 60

application of recursive filters to variational statistical analysis. Part II: Spatially 61

inhomogeneous and anisotropic general covariances. Mon. Wea. Rev., 131, pp. 1536-62

1548. 63

Riishøjgaard, L.P., R. Atlas, and G. D. Emmitt, 2008: The Impact of ATCTL 64

Observations on a Single-Level Meteorological Analysis. J. Appl. Meteorol. 44, 65

1276–1277 66

25

Salonen,K.,J.Cotton,N.Bormann,andM.Forsythe,2015:CharacterizingAMV67

Height-AssignmentErrorbyComparingBest-FitPressureStatisticsfromthe68

MetOfficeandECMWFDataAssimilationSystems.J.ApplClimatol.,54,225-24269

Santek,D.2010:TheImpactofSatellite-DerivedPolarWindsonLower-Latitude70

Forecasts.Mon.Wea.Rev.,138,123–13971

Stoffelen,A.,G.J.Marseille,E.AnderssonandD.G.H.Tan,2005:CommentsonThe72

ImpactofDopplerWindObservationsonaSingle-LevelMeteorologicalAnalysis73

byL.P.Riishojgaard,R.AtlasandG.D.Emitt,J.Appl.Meteor.,44,1276-1277.74

Su,X.,J.Derber,andJ.Jung,2012:RecentworkonSatelliteatmosphericmotion75

vectorsintheNCEPdataassimilationsystem.11thInternationalwinds76

workshop.February20-24,2012,Auckland,NewZealand77

Wu,W.-S.,R.J.PurserandD.F.Parrish,2002:Three-dimensionalvariationalanalysis78

withspatiallyinhomogeneouscovariances.Mon.Wea.Rev.,130,2905-2916.79

Yamashita,K,(2014)TheImpactofNASATERRAMISRAtmosphericMotionVector80

AssimilationintoJMAsOperationalGlobalNWPSystem.CAS/JSCWGNERes.81

Activ.Atmos.OceanicModel:http://www.wcrp-82

climate.org/WGNE/BlueBook/2014/individual-83

articles/01_Yamashita_Koji_WGNE_BB2014_MISR_yamashita_final.pdf84

Zhu, Y., L. P. Riishojgaard, R. Davies, and C. Moroney, 2004: "Assessment and 85

application of MISR winds". 7th Int. Winds Workshop, Helsinki, Finland. 86

Zhu,Y.andR.Gelaro,2008:ObservationSensitivityCalculationsUsingtheAdjointof87

theGridpointStatisticalInterpolation(GSI)AnalysisSystem.Mon.Wea.Rev.,88

26

136,335–351.doi:http://dx.doi.org/10.1175/MWR3525.189

90

27

FigureCaptions91Figure1:VerticalprofilesofpercomponentforecastimpactofUV,ATCT,andCTexperiments92

Verticalprofilesforthesouthernhemisphereextra-tropics(left;a,d,g),thetropics93(middle;b,e,h),andthenorthernhemisphereextra-tropics(right;c,f,i),areshown94forsampling(top;a,b,c),observationminus6-hourforecast(middle;d,e,f),and95forecastimpact(bottom;g,h,i)fortheuandvcomponentsofMISRCMVintheUV96experiment(labeleduvuanduvvinlegend);thealong-track(ATCTat)andcross-97track(ATCTct)inATCT;andthecross-trackcomponents(ctct)inCT.9899100Figure2:ForecastimpactofvariousobservationtypesinATCTandATCT_15experiments101

Themean24-hourforecastglobal(top;a,b)andaselectregional(bottom;c,d)102impactforselectedtypesofobservationsintheATCT_15experimentsas103accumulatedper6-hours(left;a,c)andperobservation(right;b,d).Errorbars104representingstandarddeviationsaregiven,alongsidepercentagesoftotalimpact.105106107Figure3:MappedcoverageofMISRCMVsrelativetootherclassesofobservation108

MappedcoverageforfiveclassesofobservationsassimilatedinATCTandATCT_15109experimentsspanningSep.-Nov.2014andMar.-May2015.Coverageismeasured110per2.5°latitude×2.5°longitudemapgridcellbythefractionofsix-hourperiods111throughoutexperimentsduringwhichoneormoreobservationswereassimilated112withinthatgridcell.113114115Figure4:AdjointforecastimpactofMISRCMVsrelativetootherclassesofobservation.116

AsinFigure2,butshowingmeanforecastimpactaccumulatedpersix-hourperiod117ineachmapgridcell.118119120Figure5:TimeseriesofMISRCMVsamplingandforecastimpactperorbitforATCTandATCT_15121

Timeseriesofforecastimpacts(top;a,b)andobservationcounts(bottom;c,d)for122MISRCMVdataduringATCTexperiment(left;a,c)fromSep.-Nov.2014and123ATCT_15experiment(right;b,d)fromMar.-May2015.Orbitswithanetnegative124(i.e.beneficial)forecastimpactareindicatedinblue,therestinred.Minimaand125maximaareshowninupperright.Arunningmeanover15orbits(i.e.~1day)is126plottedinblack.Numbersofobservationsperorbitthatwereassimilated(blue)127andrejected(red)areshownalongsidea15orbitrunningmean(black).128129Figure6:TimeseriesofMISRCMVsamplingandforecastimpactperorbitforNRT130

AsinFigure5,butforexperimentNRT.131132133

28

134135 136

29

137

7 Tables138

Table1Listofexperiments,durations,andassimilationmethods139

Experiment Durationofexperiment

MISRdataproduct MethodofassimilatingMISRwinds

CONTROL 2014/09/02-2014/12/01

none none

UV 2014/09/02-2014/12/01

MISRL3CMV jointuandvcomponents

ATCT 2014/09/02-2014/12/01

MISRL3CMV independentalong-trackandcross-trackcomponents

CT 2014/09/02-2014/12/01

MISRL3CMV Onlycross-trackcomponent

ATCT_15 2015/03/01-2015/06/01

MISRL3CMV independentalong-trackandcross-trackcomponents

ATCT_NRT 2014/09/02-2014/12/01

MISRNRTCMV independentalong-trackandcross-trackcomponents

140Table2Overviewofexperimentstatistics141

Label MISRobs.per6-hours

MISRobs.pervalidorbit

MISRrejectobs.pervalidorbit

MISRimpactper6-hours(Jkg-1×10-3)

MISRimpactpervalidorbit(Jkg-1×10-3)

MISRimpactperobs.(Jkg-1×10-6)

UV 6000±3000 2400±360 520±160 -12±18 -4.8±10.4 -2.0±82.5ATCT 6700±3100 2500±360 330±90 -25±18 -9.5±8.6 -3.7±78.8CT 3300±1600 1300±180 170±40 -15±15 -5.9±8.3 -4.6±107.6ATCT_15 8200±1900 2500±350 320±130 -27±17 -8.2±8.2 -3.3±79.5ATCT_NRT 6000±2200 1900±600 240±120 -21±15 -6.8±7.6 -3.5±78.1142

30

8 Figures1

Figure1:VerticalprofilesofpercomponentforecastimpactofUV,ATCT,andCTexperiments2

Verticalprofilesforthesouthernhemisphereextra-tropics(left;a,d,g),thetropics3(middle;b,e,h),andthenorthernhemisphereextra-tropics(right;c,f,i),areshown4forsampling(top;a,b,c),observationminus6-hourforecast(middle;d,e,f),and5forecastimpact(bottom;g,h,i)fortheuandvcomponentsofMISRCMVintheUV6experiment(labeleduvuanduvvinlegend);thealong-track(ATCTat)andcross-7track(ATCTct)inATCT;andthecross-trackcomponents(ctct)inCT.8

9

31

Figure2:ForecastimpactofvariousobservationtypesinATCTandATCT_15experiments10

Themean24-hourforecastglobal(top;a,b)andaselectregional(bottom;c,d)11impactforselectedtypesofobservationsintheATCT_15experimentsas12accumulatedper6-hours(left;a,c)andperobservation(right;b,d).Errorbars13representingstandarddeviationsaregiven,alongsidepercentagesoftotalimpact.1415

16Figure3:MappedcoverageofMISRCMVsrelativetootherclassesofobservation17

MappedcoverageforfiveclassesofobservationsassimilatedinATCTandATCT_1518experimentsspanningSep.-Nov.2014andMar.-May2015.Coverageismeasured19per2.5°latitude×2.5°longitudemapgridcellbythefractionofsix-hourperiods20throughoutexperimentsduringwhichoneormoreobservationswereassimilated21withinthatgridcell.2223

32

24 25

33

Figure4:AdjointforecastimpactofMISRCMVsrelativetootherclassesofobservation.26

AsinFigure2,butshowingmeanforecastimpactaccumulatedpersix-hourperiod27ineachmapgridcell.2829

30 31

34

Figure5:TimeseriesofMISRCMVsamplingandforecastimpactperorbitforATCTandATCT_1532

Timeseriesofforecastimpacts(top;a,b)andobservationcounts(bottom;c,d)for33MISRCMVdataduringATCTexperiment(left;a,c)fromSep.-Nov.2014and34ATCT_15experiment(right;b,d)fromMar.-May2015.Orbitswithanetnegative35(i.e.beneficial)forecastimpactareindicatedinblue,therestinred.Minimaand36maximaareshowninupperright.Arunningmeanover15orbits(i.e.~1day)is37plottedinblack.Numbersofobservationsperorbitthatwereassimilated(blue)38andrejected(red)areshownalongsidea15orbitrunningmean(black).3940

4142Figure6:TimeseriesofMISRCMVsamplingandforecastimpactperorbitforNRT43

AsinFigure5,butforexperimentNRT.44

4546