Data assimilation in Aladin

September 19-23 , 2005COSMO D.A. Workshop and 7th COSMO General Meeting

Data assimilation in Aladin

Data assimilation in AladinData assimilation in Aladin

Claude Fischer; Thibaut Montmerle; Ludovic Auger; Loïk Berre;

Gergely Boloni; Zahra Sahlaoui; Simona Stefanescu;Arome collaborators



Data assimilation inside AladinData assimilation inside Aladin

• 3 « operating Centers »: Budapest, Toulouse (operational), Casablanca (run daily)

• Dispatched research collaborations: Bratislava (radar), Brussels (wavelet), Bucarest (ensembles, methods), Lisbon (ensembles), Ljubljana (nudging), Prague, Tunis (surface analysis), Sofia (surface analysis and snow), Zagreb (methods)

• Existing tools: – Optimal Interpolation (« CANARI ») both for 3D and surface, – 3D-VAR (screening and minimisation), – TL and adjoint Eulerian hydrostatic models



Variational system: what is in it ?Variational system: what is in it ?

• Incremental 3D-VAR, using 80-90 % of the common Arpège/IFS code

• Continuous assimilation cycle, 6 hour frequency, long cut-off assimilation cycle and short cut-off production, coupled with Arpège, Analysis=Model gridmesh=9.5km

• Observations:– Surface pressure, SHIP winds, synop T2m and RH2m– Aircraft data– SATOB motion winds– Drifting buoys– Soundings (TEMP, PILOT)– Satellite radiances: AMSU-A, AMSU-B, HIRS, Meteosat-8 SEVIRI– No QuikSCAT

• Digital filter initialisation



Ensemble B matrixEnsemble B matrix

• Bi-periodic spectral structure functions (a torus)• Control variable: vorticity + unbalanced divergence, (T, Ps)

and specific humidity• Background error covariances are sampled from an

ensemble of Aladin forecasts, with initial conditions from an ensemble of Arpège analyses: « ensemble Jb »

• Sample: over 48 days, two pairs of 6 hour forecasts are extracted from the ensemble, and their difference is computed => 96 elements in the sample

• For the time being, constant uniform σb

bxxx

xbxHyRTxbxHyxBTxxJ

)H)((1)H)((1)(2



Data processing :• 1 pixel out of 5 is extracted and thinning boxes of 70 km are applied• IR 3.9 and 13.4 as well as ozone 9.7 are blacklisted, as well as IR channels over land• predictor, situation-dependent bias correction applied to the other channels• empirical o are used• first-guess quality control removes too large innovations• CMS/Lannion cloud classification is used to keep IR 8.7 10.8 ,12 only in clear sky while the two WV channels are also kept over low clouds

Use of Météosat-8/SEVIRI in the 3DVar of Use of Météosat-8/SEVIRI in the 3DVar of ALADINALADIN

Cloud types

Vertical weighting functions


Data assimilation in AladinDyn. Adapt. Raingauges

3DVar 3DVar with SEVIRI

Impact study :Precipitation forecast

2004/07/18 12UTCRR P12 – P6



Operational implementation in Aladin-FranceOperational implementation in Aladin-France

• A one month period over June 2003• A two week period in July 2004• First E-suite: March 23rd through May 22nd, 2005• Second E-suite: June 2nd through July 25th, 2005

=> operations started on Monday, July 25th



About scores: About scores: 3D-VAR3D-VAR versus versus Dyn. Adapt.Dyn. Adapt.

Mean sea level pressure 3 hour cumulated

precipitations2m relative humidity

bias

RMS



Scores w/r to radiosondesScores w/r to radiosondes

Temperature biasTemperature RMS

Wind biasWind RMS

Blue: 3D-VAR > Dyn Adapt

Red: 3D-VAR < Dyn Adapt


Data assimilation in AladinA specific case: June 21st, 2005



June 21st: Aladin modelJune 21st: Aladin modelP12-P6 RR

3D-VAR



June 21st: Aladin model (.)June 21st: Aladin model (.)P18-P12 RR

3D-VAR



E-suites: V1 and V2E-suites: V1 and V2• V1: 23 March through 22 May, 2005 =>

– Deteriorated MSLP bias and RMS by about 0.2 hPa– Too strong precipitation amounts at short range (6h, 12h), with significant spin-down– Analyses too wet compared to RS– No increase in the number of « Aladinades » (which is a relief, given the problems on

humidity and RR)• V2: started 2nd of June, scheduled for about 1.5 month

– Improved SEVIRI bias correction, using a case/location-dependent b.c. (with 4 predictors)

– Infrared channels over land blacklisted (problem of poor quality surface temperature)– Additional surface observations ready: T2m, RH2m => quite complementary with

SEVIRI data– Digital filter initialization: back to settings from dynamical adaptation– Reduced weighting of Jo with respect to Jb: Sb decreased from 3.24 to 2.25



Lessons from Aladin-France:Lessons from Aladin-France:• Precipitations:

– [0,3h] => caution (impact of initialisation, of imbalances …); – [3,12h] => the assimilation cycle(s) produce their own solution; – [12,24h] => a limit of predictability somewhere in this range, mostly

due to the growing influence of lateral boundary conditions; – RMS(Dyn Adap – 3DVAR) for 6 hour precipitation (mm/6h), for three

days of the test period:

07/07 08/07 22/07

P12-P6 2.32 2.08 1.75

P24-P18 1.37 1.44 1.10



3DVAR at the Hungarian Met Service3DVAR at the Hungarian Met ServiceShort history:• Implementation with French and Slovak help (June 2000)• November 2002: Quasi-operational parallel suite• Operational application (May 2005)

Basic characteristics:

• 6h 3DVAR assim. cycle

• 48h production

• linear grid

• dx ~ 8km

• 49 vertical levels



3DVAR at the Hungarian Met Service3DVAR at the Hungarian Met Service

Assimilation details:

• 6h cycle (4 long + 2 short cut-off analyses per day)

• ARPEGE fields for surface

• local 3DVAR for upper air

• NMC background error statistics

• DFI initialization

• 3h coupling in cycle (by the long cut-off ARPEGE analyses and the corresponding 3h forecasts)

Input observations:

• SYNOP: surface pressure

• TEMP: temperature, wind, pressure, specific humidity

• ATOVS/AMSU-A radiances

• AMDAR aircraft reports: temperature, wind

All the observation types above are used in the ARPEGE assimilation system too, but in a somewhat worse resolution!



3DVAR at the Hungarian Met Service3DVAR at the Hungarian Met Service

Objective scores:

• generally small improvement for temperature and wind

• neutral impact/improvement on high level’s geopotential

• degradation in low level’s geopotential and MSLP BIAS

• mixed impact on humidity

Subjective evaluation:

• improvement in T2m (0-24h)

• improvement in precipitation (0-48h)

• degradation (0-24h) / neutral impact (24-48h) in cloudiness

• neutral impact on wind

Assimilation results vs the spin-up version





In the near future, for Aladin-FranceIn the near future, for Aladin-France• 3D-VAR FGAT• Jk extra cost function to fit towards the ARPEGE analysis• Test 3-hourly analyses• Better understand the intrication between digital filter

initialization, coupling and B-matrix dynamical balances• Innovating observations for the mesoscale: sampling of

satellite data, QuikSCAT• Application for AMMA

• Code convergence with the Hirlam group: mesoscale multi-incremental 4D-VAR and an increasing collaboration on very high resolution (Arome-oriented 2.5 km)

• Next 4 year mid-term Aladin scientific plan

mid-term issues, for the collaboration:mid-term issues, for the collaboration:



4-year mid-term scientific plan: Aladin (with 4-year mid-term scientific plan: Aladin (with Arome input)Arome input)

• Variational assimilation:– « B »: wavelets, ensemble sampling, non-linear and Ω-equation,

gridpoint σb and cycling, humidity control variable– Algorithms: 3D-VAR FGAT, 4D-VAR in a nutshell– Observations: satellite (cloudy IR, microwave, locally received

data, impact of bias correction, wind retrievals), GPS zenital delay, radar reflectivity and Doppler winds, 2m and 10m mesonet

• Surface analysis:– Ocean and sea ice SAF SST derived from geostationnary

satellites– Snow analysis (with Hirlam), possibly using land SAF snow

cover– Off-line soil analysis using « dynamical OI »

• Diagnostic fine-grid 3D-VAR hourly analysis for nowcasting



Thank you for your attentionThank you for your attention



End of talkEnd of talk

Now come some extra slides for whatever …

September 19-23 , 2005 23COSMO D.A. Workshop and 7th COSMO General Meeting


Tuning of background and observation error Tuning of background and observation error variancesvariances

• Desroziers and Ivanov, 2001:(So, Sb) ?

– For a consistent systemSo = Sb = 1

• For a LAM system: – How to compute the Trace ? -> Monte-

Carlo method (Girard, 1987)– How to compute the statistical

expectation ? -> use a time mean over a suitably long range

– Samples of small size -> aggregate analysis times together

– Applied to the NMC statistics

)()(2)(

)()(2)(

min

,

min

,

KHTrJES

HKITrJES

bSSb

p

oSSo

bo

bo

Sadiki and Fischer, Tellus, 2005

Chapnik etal., QJRMS, 2004



Sensitivity experimentsSensitivity experiments

3d-Var ALADIN experiments over 1

month

NMC-laggedbackground=P06H

0.86 < 1 1.67 > 1

Standard NMC statistics

0.6 0.53

Ensemble statistics --- 1.44 (by comparison with NMC)

oS bS

18.0 oS 3d-Var ARPEGE


Data assimilation in AladinStatistics in the space of observationsStatistics in the space of observations

ARPEGE4D-VAR

ALADIN3D-VAR

radiosondes

ATOVS channels

Documents

Data assimilation in Aladin