Assimilation of GPM/DPR in Km-scale Hybrid-4DVar system

Transcription

1 Assimilation of GPM/DPR in Km-scale Hybrid-4DVar system Yasutaka Ikuta Numerical Prediction Division / Japan Meteorological Agency 6th International Symposium on Data Assimilation, 5-9 March 208, Munich, Germany

2 Operational Meso-scale NWP system in JMA Main purpose Providing disaster prevention information and aviation weather forecast Global NWP System Global Spectral Model (GSM) Horizontal resolution:tl959(0.875 deg) Global Analysis (GA): 4D-Var Meso-Scale model (MSM) and Local Forecast Model (LFM) 3-ice 6-class cloud microphysics process scheme Prognostic hydrometeors Water vapor, cloud, rain, ice, snow and graupel These hydrometeor profiles are needed to calculate reflectivity and radiance in data assimilation. Meso-Scale NWP System Meso-scale model (MSM) Horizontal resolution: 5 km Meso-Scale Analysis (MA): 4D-Var Local NWP System Local Forecast model (LFM) Horizontal resolution: 2 km Local Analysis (LA):3D-Var Analysis cycle Meso-Scale Analysis (MA) 4D-Var data assimilation system Hydrometeors are not control variables. Future Plan: Hybrid-4DVar with control variable of hydrometeors 2

3 Current state of GPM/DPR assimilation at JMA In the operational Meso-Scale NWP system Traditional 4DVAR Climatological background error covariance Hydrometeors are not control variables. GPM/DPR assimilation method Indirect assimilation (D+4DVAR) Assimilation of relative humidity profile retrieved from reflectivity Operational assimilation of GPM/DPR started in March

4 Further Utilization of GPM/DPR Characteristics of DPR 3-D information of hydrometeors Sensitivity to snow particles Covering not only land area, but also sea area can can not not be be utilized utilized in in the the current current system system Benefit of the utilization of DPR in DA Improvement of hydrometeors reproducibility in initial time -> Improvement of precipitation forecast Useful way to achieve the goal for MSM purpose 4

5 Next step of DPR assimilation Indirect assimilation using retrieved RH profiles in traditional 4DVAR Direct assimilation using reflectivity (KuPR, KaPR) profiles in new Hybrid-4DVAR 5

6 Formulation of Flow-dependent Assimilation Cost function J n2 xtn 0 B NMC x n 0 e2 xte0 B ENS x e T HM t x 0 d t R t HM t x 0 d t 2 t. Observation term 2. Climatological term 3. Flow-dependent term BG covariance is given by statistics of forecast error. BG covariance is estimated from ensemble forecasts. This term measures the fit based on the flow-dependent background error in various meteorological situations. 6

7 -moment Cloud microphysics optimized for DA Cloud particle Ice particle PSD: mono-dispersion Mass mc c D 3 kg Snow particle T=-0, Qs=0.2 g kg- Ice <--> large snow flake 6 Rain and graupel particle PSD: exponential Mass mr, g r, g D 3 kg Heymsfield and Iaquinta (2000) Thompson et al. (2008) PSD: mono-dispersion PSD: exponential + modified gamma Mass Mass 6 mi D 2 kg For midlatitude cloud (Field et al. 2007) ms D 2 kg Bulk snow density TL/AD TL/AD of of cloud cloud microphysics microphysics are are needed needed for for reflectivity reflectivity and and cloudy cloudy radiance radiance direct direct assimilation. assimilation. s.3 0 D kg m 3 7

8 Comparison of Simulation and Observation Simulation Horizontal resolution: 2 km, Lead time: 9-hour Shape: non-spherical, Snow-PSD: exponential+gamma Observation (Himawari-8 AHI) 0.4 μm Relaxation area 8

9 Test Case using Single Column Model Single column TL/AD model based on KiD* Variables P(pressure), T(temperature), Qv(specific humidity), Qx(mixing ratio of hydrometeors x) x=cloud, rain, ice, snow and graupel Vertical resolution / number of level 60 m / 00 levels Time step Initial condition 5 sec initial perturbation * Shipway and Hill (202) Forcing Vertical velocity time step 9

10 Tangent Linearization of Microphysics Scheme M t x0 Qc <- time step 2-hour -> Qr Qg Qs Qi time step time step time step time step time step time step time step time step time step time step time step time step M t x 0 x M t x 0 time step M t x time step Fixed Fixedvariables variableson onbackground backgroundtrajectory trajectory Air Airdensity densityand andpressure pressure Diffusivity Diffusivityof ofwater watervapor, vapor, thermal thermalconductivity conductivityof ofair air Nucleation Nucleation isis aa strongly strongly nonlinear nonlinear process. process. TL is almost consistent. Snow Snow error error isis originated originated from from difference difference in in ice ice perturbation perturbation 0

11 Enhancement of space-borne radar simulator Space-borne radar simulator Simplified because of reducing computation cost Beam: Not-bending Ignoring the slant beam path and beam width Effective particle: Rain, snow and graupel Ignoring cloud water and cloud ice particles Size distribution rain and graupel: Exponential distribution snow: Exponential + modified gamma distribution 20 km Particle type Spherical or Non-spherical particle Scattering calculation LUT ( Lorenz-Mie, DDA) Single scattering

12 Shape of snow particle for scattering Spherical snow Scattering: Lorentz-Mie In current method, snow particles are assumed as spherical particle. Type-A snowflakes in Liu(2008) Scattering: DDA(Discrete Dipole Approximation) These dipoles by DDSCAT 7.3 2

13 Radar Reflectivity of Snow KuPR(3.6GHz) Spherical snow particles PSD: exponential + modified gamma (Field et al. 2007) Bulk density: s.3 0 D kg m 3 Sector snowflakes KaPR(35.5GHz) SCT Liu(2008) Small Large 3

14 CFADs of radar reflectivity Contour frequency by altitude diagrams (CFADs) KuPR Observation Simulation(SCT) SCT: sector of snowflakes SPH: spherical particle of snow Simulation(SPH) KaPR Observation Simulation(SCT) Simulation(SPH) SCT is weaker than SPH, however model bias is larger than difference of particle shape. 4

15 Tangent-linear and Adjoint of Radar Simulator Function of reflectivity calculation Z Q Q, T T Z Q, T Z Z Q T Q T Q Ζ T : Tangent-linear operator Ζ T : Adjoint operator Ζ Similarity between Z Q, T and Ζ. Q Ζ Z Q Q, T T Z Q, T T KuPR (snow) KaPR (snow) Q Q 0 3, T T 0 3 Error increases in low temperature and low mass. (Maximum error 3.e-5 dbz) 5

16 Formulation of Flow-dependent Assimilation Cost function J n2 xtn 0 B NMC x n 0 e2 xte0 B ENS x e T HM t x 0 d t R t HM t x 0 d t 2 t. Observation term 2. Climatological term 3. Flow-dependent term BG covariance is given by statistics of forecast error. BG covariance is estimated from ensemble forecasts. This term measures the fit based on the flow-dependent background error in various meteorological situations. 6

17 Statistical BG Error Vertical Covariance B NMC x 6 h x 3h, x 6 h x 3h T Correlation in Jul ( sample # 8x3 forecasts ) u v w p qv qc qr qi qs q g Covariance q g q g qs qi qr qc qv p w v u qs qi qr qc qv p w v u Correlation in Jan ( sample # 8x3 forecasts ) u v w p qv qc qr qi qs q g * Cov are rescaled 7

18 Statistical BG Error Vertical Covariance Dependence on and Dependence on seasons seasons and meteorological meteorological situations situations B NMC x 6 h x 3h, x 6 h x 3h T Correlation in Jul ( sample # 8x3 forecasts ) u v w p qv qc qr qi qs q g -> -> We We need need flow-dependent flow-dependent BG BG Error Error covariance. covariance. Covariance Correlation in Jan ( sample # 8x3 forecasts ) hydrometeors u v w p qv qc qr qi qs q g q g q g qs qi qr qc qv p w v u qs qi qr qc qv p w v u hydrometeors Correlation Correlation with with hydrometeors hydrometeors hydrometeors Large Large correlation correlation * Cov are rescaled 8

19 Simplification of Background Error Covariance u Analysis variables xt u, v, Tg, ps,, Wg, p, qc, qr, qi, qs, qg Additional variables in this study Rank reduction Horizontal velocity Weak correlations between δu and δv Vertical velocity Effect is lost on early time. Pressure Substituted with hydrostatic pressure u v Tg ps Wg p qc qr qi qs q g v Tg ps Wg p qc qr qi qs q g Bu Bv Bt B Bq 9

20 Formulation of Flow-dependent Assimilation Cost function J n2 xtn 0 B NMC x n 0 e2 xte0 B ENS x e T HM t x 0 d t R t HM t x 0 d t 2 t. Observation term 2. Climatological term 3. Flow-dependent term BG covariance is given by statistics of forecast error. Extended Extended control control variable variable method method (Lorenc (Lorenc 2003) 2003) BG covariance is estimated from ensemble forecasts. This term measures the fit based on the flow-dependent background error in various meteorological situations. 20

21 Impact Experiment using Real Observation Data GPM Near-realtime Monitor ( 206/09/9 00:49:00 206/09/9 00:46:30 2

22 Impact Experiment using Real Observation Data GPM/DPR Corrected Ze (L2 Standard product) Liquid-phase: Use Solid-phase: Reject Super Observation Averaging area: 5x5 km2 Thinning Vertical interval: 500 m Now Now investigating investigating.. Bias Bias correction correction is is needed. needed. KuPR(35.5GHz) KaPR(3.6GHz) after QC and thinning 22

23 Analysis Increments Traditional DA Climatological(00%)+Ensemble(0%) Hybrid DA Climatological(50%)+Ensemble(50%) Nm: Nm: Grid Grid spacing: spacing: 55 km km Assimilation Assimilation window: window: 3-hour 3-hour Color shade: Total Precipitable Water Barbs: Surface Wind Localization radius was set to 300 km. 23

24 Traditional DA v.s Hybrid DA CNTL Forecast from Traditional DA Dry Dry Forecast from Hybrid DA Wet Wet TEST Forecast Grid spacing: 2 km TEST CNTL Color shade: Total Precipitable Water Barbs: Surface Wind 24

25 Precipitation of Forecast Traditional DA Climatological(00%)+Ensemble(0%) Observation Lead time: 9-hour 25

26 Precipitation of Forecast Hybrid DA Climatological(50%)+Ensemble(50%) Observation Lead time: 9-hour 26

27 Fractions Skill Score CNTL: Traditional DA TEST: Hybrid DA Improvement Improvement of of heavy heavy rain rain rate rate perfect skill Under Under estimation estimation of of weak weak rain rain on on largescale largescale 27

28 Summary Development of Hybrid 4D-Var Development of space-borne radar simulator with TL/AD TL/AD of cloud-microphysics scheme The TL model grows the perturbation smoothly in almost processes. Benefit of flow-dependent DA Analysis increments based of flow-dependent Improvement of precipitation forecast Reflectivity of non-spherical particles Model bias is larger than difference of scattering method. Future work Enhancement of quality control and bias correction for solid particles. More case studies targeting severe weather event. Statistical verification of long-term. 28

29 THANK YOU FOR YOUR ATTENTION 29

30 30

31 Infrared imager 0.6 μm Comparison of Simulation and Observation Shape: non-spherical, Snow-PSD: exponentioal+gamma Observation (Himwari-8 AHI) Lead time: 9-hour F05 Relaxation area Similar brightness 3

32 Infrared imager 0.6 μm Comparison of Simulation and Observation Shape: non-spherical, Snow-PSD: exponentioal+gamma Observation (Himwari-8 AHI) Lead time: 9-hour F07M Relaxation area Similar brightness 32

33 Function of AD Model of Cloud Microphysics in Identical Twin Reflectivity (KuPR) Guess True-Guess : Pseudo Observation points Brightness temperature (GMI) Guess True-Guess RTTOVSCATT (rttov.3) Satellite: GPM, Sensor: GMI, Frequency: ) 0.65 GHz(V), 2) 0.65 GHz(H), 3) 8.7 GHz(V), 4) 8.7 GHz(H), 5) 23.8 GHz, 6) 36.5 GHz(V), 7) 36.5 GHz(H), 8) 89.0 GHz(V), 9) 89.0 GHz(H), 0) 65.5 GHz(V), ) 65.5 GHz(H), 2) 83.3±3 GHz, 3) 83.3±8 GHz 33

34 Gradient Vector from pseudo observation Reflectivity Qv Qc Observation Brightness temperature Observation Saturated layer Gradient of water vapor, cloud and ice are given by AD model. Qr Qi Qs Qg Adjoint of solid-phase processes propagate the observation information to upper atmosphere. Gradient vector M T H T R HM x 34

35 Gradient Vector from pseudo observation Reflectivity Qv Qc Qr Qi Qs Qg Observation Brightness temperature Observation Saturated layer Gradient of water vapor, cloud and ice are given by AD model. AD of Microphysics Processes with ice and water vapor. Autoconversion_AD: δqs, δqi Nucleation_AD: δqv, δqi Evap/Dep_AD: δqv, δqr, δqi, δqs, δqg Q Q Melting_AD: δqr, δqi eg. Qi MTCN Qi s s Freezing_AD: δqi, δqc Collection_AD: δqc, δqi, δqs, δqg Adjoint of solid-phase processes propagate the observation information to upper atmosphere. Gradient vector M T H T R HM x 35

36 2. Enhancement of quality control and space-borne radar simulator SCATDB (Liu 2008, Honeyager et al. 206) Database of scattering coefficients for non-spherical particle. 3GHz 36GHz Liu (2008) 36

37 Flow-dependent Term Extended control variable method (Lorenc 2003) The flow-dependent method is employed extended control variables method. Background error covariance Mixing the climatological BG error and the ensemble estimated BG error Preconditioning Extended control variables Control variables /2 /2 x b0 n BNMC n e BENS e /2 BENS Cost function diag x e L/ 2,, diag x en L/ 2 N J nt n et e J o 2 2 The analysis variable is given by control variable and extended control variables. Spatial localization 37