Embed Size (px)
Citation preview
Assimilation of the GSMaP Precipitation Data with the SCALE-LETKF System
Cheng Da1 and Guo-Yuan Lien2
1 University of Maryland, College Park, USA
AICS internship report, Sep 26, 2017 1
2 RIKEN Advanced Institute for Computational Science, Kobe, Japan
Ø SCALE-LETKF
Ø Precipita4onassimila4on(Lienetal.2013,2016a,b,Kotsukietal.2017)
SCALEmodel(Nishizawaetal.2015;Satoetal.2015)
l ScalableCompu4ngforAdvancedLibraryandEnvironment(SCALE).
l Anopen-sourcebasiclibraryforweatherandclimatesimula4on.
l DevelopedbyComputa4onalClimateScienceResearchTeam.
l SCALE-RM:AregionalNWPmodelbasedonSCALE.
hUp://scale.aics.riken.jp/
SCALE-LETKF(Lienetal.2017)
hUps://github.com/gylien/scale-letkf
l ALocalEnsembleTransformKalmanFilter(LETKF)dataassimila4onpackagewiththeSCALE-RM.
l TheSCALE-LETKFis:l Aconfigurableandscalableregionaldataassimila4onsystem.l FeaturedfortheLETKFscheme,whichisnotfocusedinothercommunitydata
assimila4oncode(suchasDARTandCommunityGSI).l Testedonaverylargesupercomputer(i.e.,theKcomputer),awareofthe
efficiencyofhighperformancecompu4ng.
Experimentalnear-real-timeSCALE-LETKF
l 18-kmresolu4on;5760x4320kmarea;50membersl 6-hourcycle;5-daydeterminis4cforecastsl Testthestabilityofthesystem.l Provideimportantguidanceontheperformanceofthemodel
andthedataassimila4onse_ngs.
Anexampleof5-dayforecasts(TyphoonNANGKA;12:00UTCJuly12,2015)
Ø SCALE-LETKF
Ø Precipita4onassimila4on(Lienetal.2013,2016a,b,Kotsukietal.2017)
Satelliteprecipitationestimatesl Inrecentyears,severalsatellite-basednear-real-4meglobalprecipita4ones4mateshavebeenavailable:l TRMMMul4satellitePrecipita4onAnalysis(TMPA)/IntegratedMul4-satellitERetrievalsforGPM(IMERG)
l GlobalSatelliteMappingofPrecipita4on(GSMaP)
hUp://sharaku.eorc.jaxa.jp/GSMaP/
Assimilationofprecipitationdata
l Paststudiesofprecipita4onassimila4onshowgoodanalyses,butthemodelforgetsaboutthechangessoonabertheassimila4onstops.l Thechangeinmoistureisnotanefficientwaytoupdatethedynamicalvariablesthatprimarilydeterminetheevolu4onoftheforecastinNWPmodels.
l Flow-dependentcovariancesbetweentheprecipita4onvariableandotherdynamicalvariablesareimportant.àEnKF
l Moredifficul4es:l Thenon-Gaussianityoftheprecipita4onvariablel Thelargemodelandobserva4onerrors.
:CDFofGaussiandistribu4on:CDFoforiginalvariable
:originalvariable(mm/6hr) :Transformedvariable(sigma)
ー:Modelー:Obs.
CDF
Step0:ObtainPDF&CDF
Originalvariable (Lienetal.2013,2016b)
Gaussiantransformation
ー:Modelー:Obs.
CDF
Originalvariable (Lienetal.2013,2016b)
Step0:ObtainPDF&CDFStep1:Compute
e.g.,y=1mm/6hr
ObsF(y)
ModelF(y)
Gaussiantransformation
:CDFofGaussiandistribu4on:CDFoforiginalvariable
:originalvariable(mm/6hr) :Transformedvariable(sigma)
ー:Modelー:Obs.
CDF
Originalvariable Transformedvariable (Lienetal.2013,2016b)
Step0:ObtainPDF&CDFStep1:Compute
e.g.,y=1mm/6hr
ObsF(y)
ModelF(y)
Gaussiantransformation
:CDFofGaussiandistribu4on:CDFoforiginalvariable
:originalvariable(mm/6hr) :Transformedvariable(sigma)
ー:Modelー:Obs.
CDF
Originalvariable Transformedvariable (Lienetal.2013,2016b)
Step0:ObtainPDF&CDFStep1:ComputeStep2:Compute
Obsȳ
Modelȳ
CDF
ObsF(y)
ModelF(y)
e.g.,y=1mm/6hr
Gaussiantransformation
:CDFofGaussiandistribu4on:CDFoforiginalvariable
:originalvariable(mm/6hr) :Transformedvariable(sigma)
Original variable
Transformed variable
Gaussiantransformation
average:sigma:skewness:kurtosis:
-0.0150.7290.4180.696
sigma
average:sigma:skewness:kurtosis:
0.0802.8376.37293.91
mm/6hr
Obs-Guess(GT)
Obs-Guess(orig)
Samplingperiod:2014110100-2014110118
woGaussian-TransformaMon wGaussian-TransformaMon
MoreGaussian
(Kotsukietal.2017,JGR-A)
Gaussiantransformation:Errordistribution
ConstructionoftheprecipitationCDF
l TheCDFofprecipita4onvariablesisempiricallydeterminedbasedonthemodel/observa4onclimatologyateachgridpoint.Itrequiresalongperiodofmodel/observa4ondata.
l Lienetal.2013,2016a,b:Usingafixed1-yearclimatology
l Kotsukietal.2017:Usingthepastmonthdatapriortotheassimila4on4me
(Spin-up) (After the spin-up) (11-month average after the spin-up period)
RAOBS:Assimilaterawinsondeobserva4onsGT:Assimilaterawinsondes+uniformlydistributedglobalprecipita4onQonly:SameasGT,butonlyupdatemoisturefieldbyprecipita4onassimila4on
(Othervariablesshowsimilarresults)
Analysis 5-dayforecast
Lienetal.2013:SPEEDYmodel/idealizedexperiments
GT
GlobalresultsSolidlines:RMSerrorsDashedlines:Biases
NT
GTczandGTbz
RaobsLog
� NotransformaMon(NT)givesverybadresults.� Logarithmic(Log)transforma4onleadstomarginalresults.
� Goodformoisture,butbadfortemperature.� GaussiantransformaMon(GTczandGTbz)leadtoclearposi4veimpacts.
5-dayforecasts
Lienetal.2016a,b:GFSmodel/TMPAobservation
Kotsukietal.2017:NICAMmodel/GSMaPobservation
Spin-upperiod
ー:CTRL:RadiosondesONLYー:TEST: Radiosondes+GSMaP/Gauge(every5x5grids)
Improved!!
Summaryoftheprecipitationassimilationstudies
l Wesuccessfullyassimilatedprecipita4onwithglobalmodels,inbothidealizedOSSEsandarealis4cmodelandobserva4ons,usingtheLETKFandtheGaussiantransforma4on.l Theimpactsareseeninbothanalysesand5-dayforecasts.
l Gaussiantransforma4onisbeneficialtotheprecipita4onassimila4on:l Gaussiantransforma4onbasedontheclimatologydoesproducemoreGaussianbackgrounderrordistribu4onofprecipita4on.
l ApplyingGaussiantransforma4onseparatelytomodel/observa4onprecipita4oncancorrectthebias.
Next…theinternstudy
l Precipita4onassimila4oninaregionalmesoscalemodel?
l Impactofprecipita4onassimila4onontyphoonanalysesandforecasts?
Assimilation of the GSMaP Precipitation Data with the SCALE-LETKF System
Cheng Da
University of Maryland, College Park Department of Atmospheric & Oceanic Science
Supervisor: Guo-Yuan Lien
SponsoredbytheAICSHRdevelopment/AICSjointresearchprogram
AICS internship report, Sep 26, 2017 22
Outline 1.MoMvaMon
3.ExperimentDesign
4.Results4.1ImpactofGSMaPobserva=ons
3.1OverviewofTyphooncases
• Adjustments of the SLP/hydrometeors • Impact on TC track & intensity forecast
4.2Sensi=vitytotheQCschemes• Total number of precipitating members • Assimilating ocean obs only
5.Summary23
3.3Precip.CDFsforGaussianTransforma=on
2.ImplementaMonofPrecip.DA
3.2NWP&LETKFseOngs
Motivation • Lien et al. (2013) showed with the SPEEDY model that assimila4ng
precipita4oncan improveboth theanalysis andmedium-range forecastforalmostallvariables.
• ConsistentposiMveresultswithrealis4cglobalmodelsandobserva4ons
• Canweapplythesamemethodologytothemesoscaleregionalmodel?
ApplytheGaussiantransforma4ontothemodel&obsprecipita4on
DirectlymodifythedynamicalvariablesthroughanEnKFsystem
ü GFS-LETKF/TMPA(Lienetal.,2016b)
ü NICAM-LETKF/GSMaP(Kotsukietal.,2017)
• Specifically,canitprovideaddi4onalbenefitstotheTyphoonforecast?
Ques=on:
24
Implementation of Precip. DA component • TheSCALE-LETKFsystem(Lienetal.,2017)doesn’thavetheprecipita4on
DAcomponentyet.Soweneedtoimplementitfirst.
25
thin_gsmap.f90 thin_imerg.f90 accu_scale.f90 ppcdf.f90
mod_scale_merge.f90
mod_h5.f90 mod_nc.f90
pp_utils.f90
Precip.PreprocessingUMliMes:forbothJAXAGSMaP&
NASAIMERG
1.obsthinning2.precip.CDFcrea4on3.GrossQC
com
mon
_pre
cip.
f90
Precip.DAcomponent:&PARAM_LETKF_PRECIP
USE_PRECIP=.true., !--PPCDF_IN-- NPPX=240, NPPY=180, NCDF=200, PPZERO_THRES=1.D-3, GAUSSTAIL_THRES=1.D-3, OPT_PPTRANS=3, OPT_PPOBSERR=3,LOG_TRANS_TINY=0.6D0, MIN_OBSERR_RAIN=0.3D0, OBSERR_RAIN_PERCENT=0.5D0, OBSERR_RAIN_LT=0.18D0, OBSERR_RAIN_GT=0.5D0,RAIN_SLOT_START=5, RAIN_SLOT_END=10, MIN_RAIN_MEMBER=75 /
Obsope
LETKF
Accumulatemodelprecipita4onateach
subdomain
SCALE-LETKFsystem
Precip.w/otransforma4onPrecip.w/LogTrans.Precip.w/GaussianTrans.
config.nml.letkf:
Obs Operators / Obs Weassimilatetheaccumulated6-hrprecipita4on:
1.Thinning:
2.AggregaMon:AccumulatehourlyGSMaPdatafor6hours(7slots)
t0 t1 t2 t3t-3 t-2 t-1GSMaP
6-hourprecip.obs
pp6hr
GSMaP = 0.5× pp−3 + ppii=−2
2
∑ + 0.5× pp326
t-6
SCALE6-hourprecipfcst.
pp6hr
SCALE = ppii=−2
3
∑
t0 t1 t2 t3t-3 t-2 t-1t-6
ObsOperatorH(x):
Obsyo:
keeponlyoneGSMaPhourlyobspermodelgridGSMaPhashigherres.(10km)thanmodelres.(36km)
Experiment Design: Overviews of Two TCs in 2015
27
NangkaChan-hom
Figures from http://agora.ex.nii.ac.jp
6hrDAcyclesstartfrom0600UTC,July9to0000UTC,July17
Experiment Settings: NWP Model ModelDomains
SCALESe]ngs D01
Resolu4on 36km
Domainsize 240×180grids,36levels(~28km)
BC GFSFNLevery6hours
Microphysics TOMITA08
Radia4on MSTRNX
CumulusPara. w/oandw/KF
28
• Weat first followed configura4ons of the SCALENRT system (w/o KF).However,theprecipita4onfieldsw/oKFat36kmresolu4onisnotsimilartotheGSMaPobs.
• Lowresolu4on,inordertogetsomeresultswithinmyinternshipperiod.
Experiment Settings: LETKF LETKFse]ngs CTL GSMaP
Ensemblesize 100
Infla4on Mul4plica4veinfla4on(1.25)RTPP(0.8)
Ver4callocaliza4onscale 0.3log(pressure)
Horizontallocaliza4onscale 400km(PrepBUFR)/250km(GSMaP)
Obsassimilated. PrepBUFR PrepBUFR+GSMaP
Thresholdforprecipita4on ≥0.001mm(6hr)-1
Precipita4onCDFs JUN20~JUL6
Heightofprecipita4onobs. 850mb
QCbymin.numberofprecip.members 75/100(75%)
29
0000UTC20June
0000UTC9July
0600UTC20June
1200UTC20June
OnlyassimilatePrepBUFR Exp.GSMaP
… 0600UTC9July
Exp.CTL
0000UTC17July
Spin-upperiods:
Construction of the Precipition CDFs
30
ModelCDFs/GSMaPCDFs
CDFforGSMaP&SCALEatonemodelgridMax:335.468;Pr(ppzero)=0.63(lessfrequent)Max:162.986;Pr(ppzero)=0.40(morefrequent)
Anexample:
0000UTC20June
0000UTC9July
0600UTC20June
1200UTC20June
OnlyassimilatePrepBUFR
Exp.GSMaP
… 0600UTC9July
0000UTC17July
Spin-upperiods:
• If fixed at the same CDF value,we canget the snapshots of the model &GSMaP precipita4on values for theen4redomain.
Comparison of the Precipitation CDFs
31
SCALEw/KF SCALEw/oKF GSMaPPr(ppzero)
Morefrequent lessfrequent
Precipita
Mon
atCDF
=0.75
Smallerprecip. Largerprecip.
InaddiMon,withKF,themaximaofmodelprecip.alsodecreases(notshownhere)
Precipita
Mon
atCDF
=0.9
32
1stcycle 9thcycleCycle#1-9
CTL_GUES CTL_ANAL
GSMaP_GUES GSMaP_ANAL
Impact of GSMaP: Adjustments of the SLP
Source Min.SLP(mb)
Background 973.8
CTL_ANA 967.4
GSMaP_ANA 960.2
JMABestTrack 960
Source Min.SLP(mb)
CTL_ANA 997.8
GSMaP_ANA 985.7
JMABestTrack 945 CHAN
-HOM
NAN
GKA
33
CTL_ANA GSMaP_ANA GMI_L2_Retrievals
Adjustments of the Hydrometeors
1800UTCJUL12CWP
IWP
1800UTCJUL12 1731UTCJUL12
34
Accumulated 1-hr Precipitation Forecast
CTL_6HR_FCST GSMaP_6HR_FCST GSMaP_OBS1200UTCJUL161200UTCJUL161200UTCJUL16
• 6-hr forecast ini4alized at 0600UTC JUL 16, and validated at1200UTC,beforeNangkamadethelandfall(1400UTC)
1200UTC0600UTC 1100UTC
35
TC Track Forecast Error
Forthefirstseveralcycles,NangkamovessouthwardintheExp.CTL,whichdoesnotoccurinExp.GSMaP.
GSMaPCTL
Thickblackline:BesttrackColorlines:120-hrforecastsini4alizedatdifferent4me(warmer->laterINIT)
Averaged TC Track Forecast Error
36
FCSTerrorofCHAN-HOM FCSTerrorofNANGKA
Solid: CTL Dashed: GSMaP
120-hr Intensity Forecast Initialized at Different Time
37
CHAN-HOM NANGKA
CTL CTL
GSMaP GSMaP
Thickblackline:BesttrackColorlines:120-hrforecastsini4alizedatdifferent4me
Averaged TC Intensity Forecast Error
38
FCSTerrorofCHAN-HOM FCSTerrorofNANGKA
Solid: CTL Dashed: GSMaP
39
Sensitivity to the QC schemes NumberofPrecipitaMngmembers(GSMaP_QC50):
2.WeassimilateGSMaPwhenatleast75%membersprecipitate(~0.75inLienetal.,2016b;~0.72inKotsukietal.,2017).Theensemblesizeinthosestudies:~40.
1.Assimila4onofGSMaPhelpstointensifyNangka.However,notstrongenough.
=>WillloosingthisQCto50furtherImprovetheforecast?
GSMaPoverland(GSMaP_OCN):1. From the perspecMve of obs: Lv3 GSMaP relies on Lv2 MW retrievals.However, the quality ofMW precipita4on over land is worse than those overocean, since precipita4on referred from the IWP from High-Frequency MWchannels.2.FromtheperspecMveofmodel:Lienetal.2016ashowthatmodelsimula4onoverlandisnotstronglycorrelatedtotheobs.
=>Whatifweonlyassimilateobsoverocean?
3.Butwehave100members.
40
Accumulated 1-hr Precipitation Forecast
CTL GSMaP
GSMaP_OBS
1200UTCJUL16
1200UTCJUL161200UTCJUL16
GSMaP_QC50 GSMaP_OCN
41
Sensitivity to the QC schemes
FCSTerrorofCHAN-HOM FCSTerrorofNANGKASolid: CTL Dashed: GSMaP Dashed: GSMaP_OCN Dashed: GSMaP_QC50
42
5-day Forecast RMSE of the Entire Domain RM
SEofR
H(%
)
Forecasthours
CTL GSMaP GSMaP_OCN GSMaP_QC50
RMSEofR
H(%
)
RHat500mb
RHat1000mbRM
SEofT
(K)
RMSEofT
(K)
Tat500mb
Tat1000mb
RMSEofU
(m/s)
RMSEofU
(m/s)
Uat500mb
Uat1000mb
Forecasthours Forecasthours
43
Summary • This study inves4gates the impact of precipita4on datawithin a regional
NWPmodel. The precipita4onDA component is implementedwithin theSCALE-LETKFsystem.
• WeconducttwocasestudiesonTCChan-hom&Nangkain2015.Inbothcases,6-hraccumulatedGSMaPdataareassimilatedwiththeGaussianTransforma4on.Thepreliminaryresultsshowthat:ü TheSLP/hydrometeoranalysiswithprecip.DAismoreclosetoobs.
ü 3-daytyphoontrackforecasterroraredecreased.
ü ByloosingtheQCofprecipita4ngmembers,andremovinglandobs,wecouldfurtherimprovetheTCtrackforecasts.
ü 5-dayRHforecastatsurfaceareimproved.
44
Limitations
• Theresolu4onofthecurrentmodelsimula4onislow.
• Bothcasesdoesnotstartfromthegenesisphase.36km->18km
Newexperimentstar4ngbeforethegenesis
• Theobsmights4llbetoodense.Checktheensemblespread
• TooaggressiveQCoftheGSMaPsoverland.FurtherrefineQCs
• Prepara4onoftheCDFsrequiresalongperiod.
Currentexperimentse]ngs:
Methodology:
• Model-grid-basedCDFmightbenotwellsuitedfortheTCapplica4ons.
