for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190...

51
Ensemble 4DVAR for the NCEP hybrid GSIEnKF data assimilation system and observation impact study with the hybrid system Xuguang Wang School of Meteorology University of Oklahoma, Norman, OK 1 University of Maryland, Weather & Chaos seminar Feb. 20, 2012 OU: Ting Lei, Govindan Kutty, Yongzuo Li, Ming Xue, Brian Holland, Terra Thompson NOAA/NSSL: Lou Wicker NOAA/ESRL: Jeff Whitaker NOAA/NCEP/EMC: Daryl Kleist, Dave Parrish, Russ Treadon, John Derber

Transcript of for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190...

Page 1: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

Ensemble 4DVAR for the NCEP hybrid GSI‐EnKF data assimilation system and observation impact study with the 

hybrid system

Xuguang WangSchool of Meteorology

University of Oklahoma, Norman, OK

1University of Maryland, Weather & Chaos seminar

Feb. 20, 2012

OU: Ting Lei, Govindan Kutty, Yongzuo Li, Ming Xue, Brian Holland, Terra ThompsonNOAA/NSSL: Lou WickerNOAA/ESRL: Jeff WhitakerNOAA/NCEP/EMC: Daryl Kleist, Dave Parrish, Russ Treadon, John Derber

Page 2: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

2

Outline

Part 1: Ensemble 4DVAR for the NCEP hybrid GSI‐EnKF data assimilation system

Part 2: Observation impact study with the hybrid DA system

Part 3: Improving high resolution TC prediction using hybrid DA

Part 4: LETKF related work

Page 3: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

3

Part 1: Ensemble 4DVAR for NCEP hybrid GSI‐EnKF data assimilation system

Page 4: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

44

Hybrid GSI‐EnKF DA system

member 1 forecast

member 2 forecast

member k forecast

control forecast GSI -ECV

EnKF

control analysis

EnKFanalysis k

EnKFanalysis 2

EnKFanalysis 1

member 1 forecast

member 2 forecast

member k forecast

data assimilationFirst guess forecast

control forecast

Ensemble covariance

member 1 analysis

member 2 analysis

member k analysis

Re-center EnK

Fanalysis ensem

bleto control analysis

Wang et al. 2011

Page 5: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

5

NCEP pre‐implementation test of hybrid DAhttp://www.emc.ncep.noaa.gov/gmb/wd20rt/experiments/prd12q3r/vsdb/

Page 6: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

• Observations (e.g., satellite) are spreading through the DA window.

• Current GSI‐EnKF hybrid is 3DVAR‐based ‐‐ temporal evolution of error covariance within DA window not considered.

• ENS4DVAR is a natural extension of the current 3D GSI‐EnKFhybrid.

• Temporal evolution of the error covariance within the assimilation window is realized through the use of ensemble perturbations (e.g., Qiu et al. 2007)

6

Ensemble 4DVAR for GSI: motivation

Lei, Wang et al. 2012

Page 7: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

''1''1

2'1

1'11

211'1

21

21

21

,

HxyHxyαCαxBx

αx

oToTT

oe JJJJ

R

K

k

ekk

1

'1

' xαxx

7

B 3DVAR static covariance; R observation error covariance; K ensemble size; C correlation matrix for ensemble covariance localization; e

kx kth ensemble perturbation;'1x 3DVAR increment; 'x total (hybrid) increment; 'oy innovation vector;

H linearized observation operator; 1 weighting coefficient for static covariance;

2 weighting coefficient for ensemble covariance; α extended control variable.

• Extended control variable method in 3D GSI hybrid (Wang 2010, MWR):

Extra term associated with extended control variable

Extra increment associated with ensemble

Ensemble 4DVAR for GSI: method

Add time dimension in ENS4DVAR

Page 8: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

8

One obs. example for TC 

‐3h                                 0                               3h*

ens4dvar ens3dvar

–3h increment propagated by model 

integration 

t=0 t=0 t=0

time

Page 9: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

9

tt‐3h t+3h

Temp.

Height

t‐3h t t+3h

Upstream impact

Downstream impact

Another example 

Page 10: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

10

Why Hybrid? “Best of both worlds” VAR (3D,4D)

EnKF hybrid References (examples)

Benefit from use of flow dependent ensemble covariance instead of static B 

x x Hamill and Snyder 2000; Lorenc 2003, Wang et al. 2007b,2008ab, 2009b; Zhang et al. 2009; Buehneret al. 2010ab; Wang 2011;

Robust for small ensemble x Wang et al. 2007b, 2009b; Buehner et al. 2010b

Better localization for integrated measure, e.g. satellite radiance; radar with attenuation

x Campbell et al. 2010

Easiness to add various constraints

x x

Outer loops  x x

More use of various existing capability in VAR

x x

Summarized in Wang 2010,  MWR

Page 11: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

11

ExperimentsTest period: Aug. 15 2010 – Sep. 20 2010

Model: GFS T190 64 levels

Observations: all operational data 

Data assimilation methods:  o GSI (gsi)o 3D GSI‐EnKF (hybrid1way)o ensemble 4DVAR: 2‐hourly frequency (ens4dvar1way) 1‐hourly frequency (ens4dvar1way‐hrly)o excluding the balance constraint: hybrid1way‐nb ens4dvar1way‐nb

Page 12: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

12

Hurricane track forecasts

2010 hurricanes 

• ens3dvar hybrid better than GSI and further improvement by ens4dvar hybrid.  • Balance constraint in GSI hurt TC forecast for both ens3dvar and ens4dvar hybrid.

Page 13: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

13

Global forecasts verified against EC analyses

Height Temperature

• ens3dvar hybrid better than GSI and further improvement by ens4dvar hybrid.  • Balance constraint in GSI help both ens3dvar and ens4dvar hybrid.

Page 14: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

1414

Global forecasts verified against conv. obs.

Significant improvement of ens3dvar hybrid and ens4dvar hybrid over GSIens4dvar showed further improvement over ens3dvar especially for wind

6h wind 6h temp

Page 15: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

1515

Global forecasts verified against conv. obs.

Significant improvement of ens3dvar hybrid and ens4dvar hybrid over GSIens4dvar showed further improvement over ens3dvar especially when “nb” balance constraint seems helpful at early lead time, but hurt at later lead time for hybrid

96h wind 96h temp

Page 16: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

Summary

Ensemble‐4DVAR capability was developed for GSI and  tested for GFS.  Ensemble‐4DVAR further improved upon the ENS3DVAR hybrid for TC track forecasts, global forecasts verified against EC and obs.  

Depending on the specific forecasts, the balance constraint in the GSI can benefit or hurt the forecasts .  Improvements on balance constraint needed.

16

Page 17: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

17

Part 2: Observation impact study with the hybrid DA system

Page 18: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

Introduction

Modern data assimilation systems assimilate millions ofobservations each day to produce initial conditions fornumerical weather forecasts.

What are the impacts of the obs. from different platforms? Are impacts dependent on DA methods and how? Are impacts dependent on how impacts are assessed and

how?

18

Page 19: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

19

Increments with static and flow dependent B

Page 20: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

Increments by GSI and hybrid: AMSU

20

Page 21: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

Introduction (cont’d)

Various methods have been used to assess the impacts ofobservations

Observing System Experiments (OSE; e.g., Zapotocny et al.,2007, etc.)

Adjoint method (e.g., Baker and Daley, 2000; Gelaro et al.,2008, etc.)

Ensemble based method (e.g., Liu and Kalnay, 2008; Kalnay2011; Kunii et al. 2012, etc.)

21

Page 22: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

Experiment design

Test period : Dec. 15 ‐ Jan.31 2010Model: Global Forecast System Model (GFS) T190 64 levels Observations  denied: AMSU, Rawinsonde Data assimilation methods: GSI Hybrid GSI‐EnKF (one way coupled) 6 experiments for OSE GSI assimilating all obs. (gsi) GSI denied rawinsonde (gsi noraob) GSI denied AMSU (gsi noamsu) Hybrid assimilating all obs. (hybrid) Hybrid denied rawinsonde (hybrid noraob) Hybrid denied AMSU (hybrid noamsu) 

22Kutty, Wang et al. 2012

Page 23: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

Experiment design (cont’d)

Ensemble based obs. impact estimate (Kalnay 2011)

Goal evaluate the quality of the estimate for the hybrid system research to further improve the estimate compare with other obs. impact metric, e.g., OSE

Initial experiments Estimate of rawinsonde temp. impact for temp. and wind forecasts. Verified w.r.t  actual impact. Same for satellite data (e.g., AMSU)

23

Page 24: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

RMSE of global forecasts w.r.t EC analysis

24h wind 24h temperature 24h spec. humidity

24

Page 25: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

RMSE of global forecasts w.r.t obs.

24h wind 24h temperature

25

Page 26: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

Zonally averaged impact (wind RMSE difference)

26

Page 27: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

Zonally averaged impact (temp. RMSE difference)

27

Page 28: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

Zonally averaged impact (humid. RMSE difference)

28

Page 29: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

Relative impact (500mb Height)

(b) gsi noraob(a) gsi noamsu

(c) hybrid noamsu (d) hybrid noraob

29

Page 30: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

1000hPa

925hPa

850hPa

700hPa

500hPa

300hPa

250hPa

200hPa

global

-0.0035 -0.0030 -0.0025 -0.0020 -0.0015 -0.0010 -0.0005 0.0000

estimate actual

error reduction / gridpoint (K2)

Ensemble based obs. impact estimate for the hybrid system

1000hPa

925hPa

850hPa

700hPa

500hPa

300hPa

250hPa

200hPa

global

-0.008 -0.006 -0.004 -0.002 0.000

estimate actual

error reduction /grid point (ms-1)2

Rawinsonde temp. impact for 24h temp. fcst

Rawinsonde temp. impact for 24h wind. fcst

30

Page 31: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

OSE Forecasts by hybrid DA was better than GSI in both control and data denial 

experiments.  In some verifications, the hybrid assimilating less data was even better than the GSI assimilating all obs.

Magnitude and distribution of obs. impact depend on the obs. and the DA methods. The relative impact between rawinsonde and AMSU depends on DA methods and verifications.

Ensemble based obs. impact estimate for the hybrid Initial results showed the estimate was promising. The quality of the estimate can be dependent on obs. and forecasts of interest, forecast lead time, etc.

Summary

31

Page 32: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

32

Part 3: Improving high resolution TC prediction using hybrid DA

Page 33: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

33

•WRF: Δx=5km

•Observations: radial velocity from two WSR88D radars (KHGX, KLCH)

Radar hybrid DA for Ike 2008

Li, Wang, Xue 2012, MWR

IKE 2008

Page 34: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

34

Temperature increment

3DVARb Hybrid1  Hybrid.5

Page 35: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

35

No Radar DA 3DVARb

Hybrid1 Hybrid.5

Cold core

Warm core

Wind and pot. temperature analyses

Page 36: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

36

Wind and Precipitation forecasts

Page 37: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

37

Part 4: LETKF related work

Effects of sequential or simultaneous assimilation and localization methods on the performance of ensemble Kalmanfilter

Doppler Radar data assimilation using the LETKF

Page 38: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

Motivation

Popular EnKF schemes contain differences(1) Stochastic versus deterministic

EnKF vs. LETKF, EnSRF, ETKF, etc.(2) Simultaneous vs. Sequential

LETKF vs. EnSRF, EnKF(3) If and how localization is applied

B vs. R localization Previous studies (e.g. Kepert 2009; Greybush et al. 2011; Janjic 2011) 

related to (2) & (3) did not reach consensus Rigorous study on whether the choice of (2) and (3) has an effect on 

EnKF accuracyo Optimally tunedo Study the effect of (2) and (3) in isolation and their interactiono Comparison in variety of contexts 

Holland and Wang, 2012, QJRMS

Page 39: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

Experimental Design

• 2‐layer dry global primitive equation model (Zou et al. 1993)

• Model error: T31 model truncation, T127 Truth run• Observations = truth run + added error• Comparison in various contexts

Baseline experiments: 50 member, Surface π &Interface height obs. Other experiments: with digital filter initialization, 

adding more observation types, reducing observation number, increasing ensemble size, decreasing Pb/R ratio.

Page 40: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

Scheme Breakdown

SchemeSerial/

simultaneous Localization Algorithm

Bserial Serial B EnSRF

Rserial Serial R EnSRF

Bsimult Simult B EnSRF

Rsimult Simult R EnSRF

LETKF Simult R LETKF

LETKF: Hunt et al. 2007

EnSRF: Whitaker and Hamill 2002

Page 41: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

Baseline experiments: Analysis Error

Page 42: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

Baseline experiments: imbalance

Page 43: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

Explore imbalance difference using non‐cycling experiments

Surface π tendency increment difference (Bsimult‐Rsimult)

Wind and height gradient increment difference (Bsimult‐Rsimult)

Page 44: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

44

Conclusion 

Baseline experiment:• Effects of sequential vs. simultaneous assimilation depend on 

whether B‐ or R‐localization was being used (Bsimult worse than Bserial; Rsimult better than Rserial)

• Effects of B‐ or R‐localization depend on simultaneous or serial assimilation was being used (Bserial ~ Rserial; Bsimult worse than Rsimult)

• Schemes that produced less accurate analyses tended to be least balanced.

Sensitivity experiments:• Difference among the schemes reduced when digital filter 

initialization was used, both wind and mass observations were assimilated,  the number of observations was decreased, the ensemble size was increased, and the ratio of forecast error to observation error in the system was decreased.

Page 45: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

Doppler Radar Data Assimilation Using LETKF

Apply LETKF to storm‐scale radar data assimilation

Study the effect of simultaneous/serial assimilation and localization methods for storm scale by comparing with serial EnSRF

Study the scalability of parallel LETKF.

Thompson, Wicker and Wang, 2012

Page 46: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

Vertical Velocity

Refle

ctivity

Truth LETKF EnSRF

Preliminary OSSE results after 1 hour of DA

Page 47: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

47

Ongoing and future work  

Working with collaborators, further develop, test and research on ens4dvar and TL/ADJ 4DVAR hybrid for Global DA.

Working with collaborators, further test and research the hybrid (ens3dvar, ens4dvar) for regional NWP.

Conduct OSE for hybrid DA for other data and other forecasts of interest. Conduct OSE using ens4dvar. Research and improve the ensemble based obs. impact estimation for different 

data types and different forecasts for the hybrid DA system. Research on obs. impacts inferred by e.g., OSE, ensemble based method and 

adjoint method. Continue efforts on LETKF for storm scale radar assimilation.

Page 48: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

48

ReferencesCampbell, W. F., C. H. Bishop, D. Hodyss, 2010: Vertical Covariance Localization for Satellite Radiances in Ensemble KalmanFilters. Mon. Wea. Rev., 282‐290.Lorenc, A. C.  2003: The potential of the ensemble Kalman filter for NWP – a comparison with 4D‐VAR. Quart. J. Roy. Meteor. Soc., 129, 3183‐3203.Buehner, M., 2005: Ensemble‐derived stationary and flow‐dependent background‐error covariances: evaluation in a quasi‐operational NWP setting. Quart. J. Roy. Meteor. Soc., 131, 1013‐1043.Hamill, T. and C. Snyder, 2000: A Hybrid Ensemble Kalman Filter–3D Variational Analysis Scheme. Mon. Wea. Rev., 128, 2905‐2915. Wang, X., C. Snyder, and T. M. Hamill, 2007a: On the theoretical equivalence of differently proposed ensemble/3D‐Var hybrid analysis schemes. Mon. Wea. Rev., 135, 222‐227. Wang, X., T. M. Hamill, J. S. Whitaker and C. H. Bishop, 2007b: A comparison of hybrid ensemble transform Kalman filter‐OI and ensemble square‐root filter analysis schemes. Mon. Wea. Rev., 135, 1055‐1076. Wang, X., D. Barker, C. Snyder, T. M. Hamill, 2008a: A hybrid ETKF‐3DVar data assimilation scheme for the WRF model. Part I: observing system simulation experiment. Mon. Wea. Rev., 136, 5116‐5131. Wang, X., D. Barker, C. Snyder, T. M. Hamill, 2008b: A hybrid ETKF‐3DVar data assimilation scheme for the WRF model. Part II: real observation experiments. Mon. Wea. Rev., 136, 5132‐5147. Wang, X., T. M. Hamill, J. S. Whitaker, C. H. Bishop, 2009: A comparison of the hybrid and EnSRF analysis schemes in the presence of model error due to unresolved scales. Mon. Wea. Rev., 137, 3219‐3232.Wang, X., 2010: Incorporating ensemble covariance in the Gridpoint Statistical Interpolation (GSI) variational minimization: a mathematical framework. Mon. Wea. Rev., 138,2990‐2995. Wang, X. 2011: Application of the WRF hybrid ETKF‐3DVAR data assimilation system for hurricane track forecasts. Wea. Forecasting, in press. Li, Y, X. Wang and M. Xue, 2011: Radar data assimilation using a hybrid ensemble‐variational analysis method for the prediction of hurricane IKE 2008. Mon. Wea. Rev., submitted.Buehner, M, P. L. Houtekamer, C. Charette, H. L. Mitchell, B. He, 2010: Intercomparison of Variational Data Assimilation and the Ensemble Kalman Filter for Global Deterministic NWP. Part I: Description and Single‐Observation Experiments. Mon. Wea. Rev., 138,1550‐1566. Buehner, M, P. L. Houtekamer, C. Charette, H. L. Mitchell, B. He, 2010: Intercomparison of Variational Data Assimilation and the Ensemble Kalman Filter for Global Deterministic NWP. Part II: One‐Month Experiments with Real Observations. Mon. Wea. Rev., 138,1550‐1566. Holland, B. and X. Wang, 2011: Effects of sequential or simultaneous assimilation of observations and localization methods on the performance of the ensemble Kalman filter . Q. J. R. Meteo. Soc., submitted

Page 49: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

49

Flow‐dependent covariance

K

kk

GSI (static covariance) Hybrid (ensemble covariance)

Wang et al. 2011

Page 50: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

50

Track and intensity forecasts

No radar

No radar

Li et al. 2012

Page 51: for the NCEP hybrid GSI EnKF data - University Of … 4DVAR for the NCEP hybrid ... Model: GFS T190 64 levels Observations: all operational data ...

51

3DVAR with un‐tuned static covariance (3DVARa)

3DVAR with tuned static covariance (3DVARb) 

Hybrid with full ensemble covariance (hybrid1)

Hybrid with half ensemble covariance and half static covariance (hybrid.5)

Wind increment

Li et al. 2012