1. christoph.schraff@dwd.de KENDA & WG1 Overview COSMO GM, Rome, 6 September 2011 1 christoph.schraff@dwd.de Deutscher Wetterdienst, D-63067 Offenbach, Germany current DA method: nudging PP KENDA for km-scale EPS: LETKF radar-derived rain rates: latent heat nudging LETKF for HRM (final setup) WG1 Overview Long-term impact of LHN assimilation on soil moisture ?

2. christoph.schraff@dwd.de KENDA & WG1 Overview COSMO GM, Rome, 6 September 2011 2 impact of LHN on soil moisture June 2010 differences in normalised soil moisture (LHN - noLHN) inside radar data domain Experiment: 18 months (April 2009 – August 2010) COSMO-DE without LHN  compare with operational COSMO-DE with LHN

3. christoph.schraff@dwd.de KENDA & WG1 Overview COSMO GM, Rome, 6 September 2011 3 impact of LHN on soil moisture: impact on surface parameters June 2010, 12 UTC runs BIAS LHN noLHN bias reduced for different parameters

4. christoph.schraff@dwd.de KENDA & WG1 Overview COSMO GM, Rome, 6 September 2011 4 model runsTd_2m SummerWinter monthly relative forecast skill (1 – rmse(LHN) / rmse(noLHN)) Summer T_2m 30%35%  LHN  higher soil moisture content  improved T-2m, Td-2m forecasts, benefit lasts over whole forecast time impact of LHN on soil moisture: impact on surface parameters March 2009Aug 2010 dew point depression

5. christoph.schraff@dwd.de KENDA & WG1 Overview COSMO GM, Rome, 6 September 2011 5 christoph.schraff@dwd.de Deutscher Wetterdienst, D-63067 Offenbach, Germany current DA method: nudging PP KENDA for km-scale EPS: LETKF Satellite radiances: 1DVAR: AMSU-A, SEVIRI, IASI (no benefit) spatial AMSU-A obs errors statistics radar-derived rain rates: latent heat nudging 1DVar (comparison) radar radial wind: VAD (at DWD: no benefit) nudg. V r (being implemented) GPS-IWV (ceased) LETKF for HRM (final setup) surface (snow …)  PP COLOBOC screen-level: scatterometer wind (OSCAT, OceanSat2) improved T2m assimilation, adapt T_soil WG1 Overview daytime biases in summer: Vaisala RS-92 humidity obs dry bias  bias correction operational model warm bias (low levels) Vadim Gorin, Mikhail Tsyrulnikov: Estimation of multivariate observation-error statistics for AMSU-A data (MWR, accepted)

6. christoph.schraff@dwd.de KENDA & WG1 Overview COSMO GM, Rome, 6 September 2011 6 Status Overview on PP KENDA Km-scale ENsemble-based Data Assimilation Contributions / input by: Hendrik Reich, Andreas Rhodin, Roland Potthast, Uli Blahak, Klaus Stephan (DWD) Annika Schomburg (DWD / Eumetat) Yuefei Zeng, Dorit Epperlein (KIT Karlsruhe) Daniel Leuenberger, Manuel Bischof (MeteoSwiss) Mikhail Tsyrulnikov, Vadim Gorin (HMC) Lucio Torrisi (CNMCA) Amalia Iriza (NMA) christoph.schraff@dwd.de Deutscher Wetterdienst, D-63067 Offenbach, Germany Task 1: General issues in the convective scale (e.g. non-Gaussianity) Task 2: Technical implementation of an ensemble DA framework / LETKF Task 3: Tackling major scientific issues, tuning, comparison with nudging Task 4: Inclusion of additional observations in LETKF

7. christoph.schraff@dwd.de KENDA & WG1 Overview COSMO GM, Rome, 6 September 2011 7 analysis step (LETKF) outside COSMO code  ensemble of independent COSMO runs up to next analysis time  separate analysis step code, LETKF included in 3DVAR code of DWD Task 2: Technical implementation of an ensemble DA framework / LETKF collect obs from ]t ana –3h, t ana ] old setup: obs from ]t ana –1.5h, t ana +1.5h] thinned

8. christoph.schraff@dwd.de KENDA & WG1 Overview COSMO GM, Rome, 6 September 2011 8 modifications in COSMO fully implemented, but not yet in official code (also in order to have a sub-hourly update frequency) first version LETKF (KENDA) implemented in NUMEX, to be tested  still use scripts to do a few stand-alone cycles with LETKF Task 2: Technical implementation of an ensemble DA framework / LETKF deterministic analysis implemented  talk of Hendrik LBC at analysis time : use (perturbed) analysis ensemble member itself instead of (temporally interpolated) LBC fields from steering ensemble system next step:LBC from GME LETKF ensemble interpolate GME ensemble perturbations and add to deterministic COSMO-DE LBC

9. christoph.schraff@dwd.de KENDA & WG1 Overview COSMO GM, Rome, 6 September 2011 9 Analysis for a deterministic forecast run : use Kalman Gain K of analysis mean L : interpolation of analysis increments from grid of LETKF ensemble to (possibly finer) grid of deterministic run deterministic ensemble deterministic analysis recently implemented Kalman gain / analysis increments not optimal, if deterministic background x B (strongly) deviates from ensemble mean background deterministic run must use same set of observations as the ensemble system !

10. christoph.schraff@dwd.de KENDA & WG1 Overview COSMO GM, Rome, 6 September 2011 10  build COSMO modules with clean interfaces & introduce them in LETKF no further progress still plan to first implement upper-air obs (needed as input for VERSUS) ‘stat’-utility: compute model (forecast) – obs for complete NEFF (NetCDF feedback files) :  adapt verification mode of 3DVar/LETKF package  need to implement COSMO observation operators with QC in 3DVar/LETKF package  hybrid 3DVar–EnKF approaches in principle applicable to COSMO ) Task 2:Technical implementation: verification for KENDA

11. christoph.schraff@dwd.de KENDA & WG1 Overview COSMO GM, Rome, 6 September 2011 11 NEFFprove (Amalia Iriza, NMA) : tool to compute and plot verification scores based on NetCDF feedback files (NEFF) computes statistical scores for different runs (‘experiments’), focus: use exactly the same observation set in each experiment !  select obs according to namelist values (area, quality + status of obs, … ) compute scores (upper-air obs, screen-level obs, RR, cloud, etc.) (for each experiment, variable, forecast time, vertical level, …) - deterministic continuous: BIAS, RMSE, MSE - dichotomous: accuracy, Heidke Skill Score, Hanssen and Kuiper discriminant - ensemble scores (to be done ): reliability, ROC, Brier Skill Score, (continuous) Ranked Probability Score plot the scores, using GNUplot (work is in progress) Task 2:Technical implementation: verification for KENDA

12. christoph.schraff@dwd.de KENDA & WG1 Overview COSMO GM, Rome, 6 September 2011 12 –lack of spread: (partly ?) due to model error and limited ensemble size which is not accounted for so far to account for it: covariance inflation, what is needed ?  multiplicative (tuning, or adaptive ( y – H(x) ~ R + H T P b H ))  additive ((e.g. statistical 3DVAR-B), stochastic physics (Torrisi)) –localisation (multi-scale data assimilation, 2 successive LETKF steps with different obs / localisation ?) Task 3:scientific issues & refinement of LETKF –model bias –update frequency :up to now only 3-hourly –non-linear aspects, convection initiation (latent heat nudging ?) –technical aspects: efficiency, system robustness (Nov. 2012: regular ‘pre-operational’ LETKF suite)  talk by Hendrik Reich

13. christoph.schraff@dwd.de KENDA & WG1 Overview COSMO GM, Rome, 6 September 2011 13 investigate LETKF in Observing System Simulation Experiments (OSSE) –apply LETKF to idealized convective weather systems, tune LETKF settings (localization, covariance inflation) –quantify + reduce spin-up time to assimilate convective storm by LETKF MeteoSwiss: plan for 2-year project not accepted LETKF scientific issues / refinements: investigate idealised convection –however master thesis (5 months): Manuel Bischof, MeteoSwiss perturbations of the storm environment for ensemble generation: –wind speed, temperature (low-level stability), and (low-level) humidity all suitable to perturb –meaningful perturbation amplitudes: 2 m/s, 0.25 K, 2 %

14. christoph.schraff@dwd.de KENDA & WG1 Overview COSMO GM, Rome, 6 September 2011 14 stochastic physics (Palmer et al., 2009) : by Lucio Torrisi, available Oct. 2011 LETKF scientific issues / refinements: model error estimation

15. christoph.schraff@dwd.de KENDA & WG1 Overview COSMO GM, Rome, 6 September 2011 15 LETKF scientific issues / refinements: model error estimation objective estimation and modelling of model (tendency) errors (not to be confused with forecast errors) M. Tsyrulnikov, V. Gorin (HMC Russia) 1.set up a (revised) stochastic model (parameterisation) for model error (ME) MEM : e =  * F phys (x) + e add – involves stochastic physics (  * F phys (x) ) and additive components e add and includes multi-variate and spatio-temporal aspects – model error model parameters: (E(  ), D(  ), D(e add )) 2.develop MEM Estimator : estimate parameters by fitting to statistics from real forecast (COSMO-RU-14) and observation (radiosonde) tendency data, using a revised (maximum likelihood based) method (tested using bootstrap) 3.develop ME Generator embedded in COSMO code 4.develop MEM Validator: exactly known ME in OSSE set-up, retrieve with MEM Estimator  The reproducibility of the MEM parameters is not yet established !

16. christoph.schraff@dwd.de KENDA & WG1 Overview COSMO GM, Rome, 6 September 2011 16 radar : assimilate 3-D radial velocity and 3-D reflectivity directly Uli Blahak (DWD), Yuefei Zeng, Dorit Epperlein (KIT Karlsruhe) 1.Implement full, sophisticated observation operators (by end 2011) –compute reflectivity from model values using Mie- or Rayleigh-scattering Task 4: use of additional observations Tasks 4.1, 4.2: use of radar obs

17. christoph.schraff@dwd.de KENDA & WG1 Overview COSMO GM, Rome, 6 September 2011 17 averaging over beam weighting function: weighted spatial mean over measuring volumn (cross-beam vertically / horizontally) propagation of radar beam : ‘online’ depend. on refractive index, or approximate standard atmosphere MPP parallelisation: beam in an azimuthal slice has to be collected on 1 PE sub-domain 1sub-domain 2 Tasks 4.1, 4.2:use of radar obs : radar observation operator

18. christoph.schraff@dwd.de KENDA & WG1 Overview COSMO GM, Rome, 6 September 2011 18 Tasks 4.1, 4.2:use of radar obs : radar observation operator without attenuationwith attenuation of radar reflectivity by atmospheric gases + hydrometeors (using the extinction coefficients)

19. christoph.schraff@dwd.de KENDA & WG1 Overview COSMO GM, Rome, 6 September 2011 19 radar : assimilate 3-D radial velocity and 3-D reflectivity directly Uli Blahak (DWD), Yuefei Zeng, Dorit Epperlein (KIT Karlsruhe) 1.implement full, sophisticated observation operators (by end 2011) 2.apply / test sufficiently accurate and efficient approximations –by looking at the simulated obs (by March 2012) –doing assimilation experiments : OSSE setup Task 4: use of additional observations Tasks 4.1, 4.2: use of radar obs

20. christoph.schraff@dwd.de KENDA & WG1 Overview COSMO GM, Rome, 6 September 2011 20 ground-based GNSS slant path delay (early 2012, N.N.) –implement non-local obs operator in parallel model environment –test and possibly compare with using GNSS data in the form of IWV or of tomographic refractivity profiles Particular issue: localisation for (vertically and horizontally) non-local obs Task 4.3: use of GNSS slant path delay

21. christoph.schraff@dwd.de KENDA & WG1 Overview COSMO GM, Rome, 6 September 2011 21 cloud information based on satellite and conventional data Annika Schomburg (DWD / Eumetsat) –derive incomplete analysis of cloud top + cloud base, using conventional obs (synop, radiosonde, ceilometer) and NWC-SAF cloud products from SEVIRI –use obs increments of cloud or cloud top / base height or derived humidity –use SEVIRI brightness temperature directly in LETKF in cloudy (+ cloud-free) conditions, in view of improving the horizontal distribution of cloud and the height of its top –compare approaches Particular issues:non-linear observation operators, non-Gaussian distribution of observation increments Task 4.4: use of cloud info

22. christoph.schraff@dwd.de KENDA & WG1 Overview COSMO GM, Rome, 6 September 2011 22 Task 4.4: use of cloud info –make observations available –aggregation / interpolation to COSMO-DE / COSMO-EU grids –regular archiving: –NWC-SAF cloud products available since May 2011: (CT: cloud type, CTT: cloud top temperature, CTH: cloud top height) –BT available since June 2011: all 12 SEVIRI channels –assumption in LETKF: no bias  look at systematic differences in BT / cloud products

23. christoph.schraff@dwd.de KENDA & WG1 Overview COSMO GM, Rome, 6 September 2011 23 high clouds: cold bias of model-BT, (partly ?) due to known bias in RTTOV-9 BT Task 4.4: use of cloud info

24. christoph.schraff@dwd.de KENDA & WG1 Overview COSMO GM, Rome, 6 September 2011 24 fractional water clds high semitransparent very highclouds highclouds mediumclouds lowclouds very lowclouds cloud-free water cloud-free land undefined cloud type CTcloud top height CTH NWC-SAF SEVIRI cloud products: example COSMO: cloud water q c > 0, or cloud ice q i > 5. 10 -5 kg/kg  clc = 100 % subgrid-scale clouds  clc = f(RH; shallow convection; q i, q i,sgs ) < 100 % Task 4.4: use of cloud info

25. christoph.schraff@dwd.de KENDA & WG1 Overview COSMO GM, Rome, 6 September 2011 25 cloud top height CTH distribution May – July 2011 clc > 50%clc > 70%clc > 90% clc > 20%clc > 1% NWC-SAF CTH ‘obs’ COSMO CTH for clc > x % Task 4.4: use of cloud info

26. christoph.schraff@dwd.de KENDA & WG1 Overview COSMO GM, Rome, 6 September 2011 26 cloud top height CTH distribution May – July 2011 NWC-SAF CTH ‘obs’ COSMO CTH with optimal threshold: for clc > 70 % for levels above 5000 m, for clc > 40 % for levels below 5000 m Task 4.4: use of cloud info

27. christoph.schraff@dwd.de KENDA & WG1 Overview COSMO GM, Rome, 6 September 2011 27 clc > 50%clc > 30%clc > 0.1%clc > 70%clc > 90% height [km] optimal clc CTH COSMO 6-h forecast, for 18 May 2011, 18 UTC cloud cover [%], shading: model isoline: CTH MSG Task 4.4: use of cloud info

28. christoph.schraff@dwd.de KENDA & WG1 Overview COSMO GM, Rome, 6 September 2011 28 Technical implementation of verification for KENDA thank you for your attention

29. christoph.schraff@dwd.de KENDA & WG1 Overview COSMO GM, Rome, 6 September 2011 29 + ( - ) 0.9 Pert 1 -0.1 Pert 2 -0.1 Pert 3 -0.1 Pert 4 flow-dep background error cov. P b = ( k – 1 ) X b ( X b ) T k perturbed forecasts ensemble mean fcst. forecast perturbations X b Local Ensemble Transform Kalman Filter LETKF (Hunt et al., 2007) ensemble mean forecast local transform matrix w analysis mean analysis error covariance (computed only in ensemble space) in the (k-1) -dimensional (!) sub-space S spanned by background perturbations : set up cost function J(w) in ensemble space, explicit solution for minimisation (Hunt et al., 2007) analysis perturbations W a(i) perturbed analyses