Towards an adaptive observation network: monitoring the observations impact in ECMWF forecast

1. Towards an adaptive observation network: monitoring the observations impact in ECMWF forecast Carla Cardinali Office 1006

2. Mid-1995, Roger Daley joined NRL data assimilation when predictability scientists were preparing the FASTEX targeting campaign. It was clear that while sensitivity gradients products could identify sensitivity regions, they could not provide any guidance on how deployed extra-observations would have improved the forecast. After one of the frequent discussions on the subject, Roger left the office wondering if there was some way to find the sensitivity of the forecast model with respect to the observations. Unable to sleep that night ,he instead derived the equations for the sensitivity of the forecast aspect to the observation and the background. Roger Daley’s idea J is a measure ofthe forecast error (as defined through e.g dry energy norm)

3. Forecast Sensitivity to Observation or Observation Impact on Forecast • Equation • FSO Diagnostic Tool • Forecast system performance investigation in two different seasons • Monitoring the forecast impact • ECMWF Operational configuration • Conclusion Outline

4. J is a measure ofthe forecast error (as defined through e.g dry energy norm) Forecast sensitivity to observation: Equationsfrom a Roger Daley idea Forecast error sensitivity to the analysis Rabier F, et al. 1996

5. Summer Sensitivity Gradient Dry Energy norm Winter + -

6. J is a measure ofthe forecast error (as defined through e.g dry energy norm) Forecast sensitivity to observation: Equationsfrom a Roger Daley idea Forecast error sensitivity to the analysis Rabier F, et al. 1996

7. xb y y xb xb y Analysis Solution in Model space B(qxq)=Var(xb) R(pxp)=Var(y) K(qxp) gain matrix H(pxq) Jacobian matrix in observation space Analysis Sensitivity to the Observation DFS

8. Define Forecast Sensitivity Analysis sensitivity to the observation (model space)

9. Equations Solution for forecast sensitivity Krylov Subspace Method

10. J is a measure ofthe forecast error (as defined through e.g dry energy norm) Forecast error sensitivity to the analysis Rabier F, et al. 1996 Forecast sensitivity to observation: Equationsfrom a Roger Daley idea Compute the forecast impact or forecast error variation δJ

11. 15 June-15 July Summer 2006 Monitoring ECMWF System 24h OSE FcE Cycle 31R2 T511T95T159 L60 5 January-12 February Winter 2007

12. Negative impact Positive impact FSO: Pilot and Wind Profilers FcE contribution Summer 2006 Pilot Wind Profiler NA

13. FSO: Wind Profilers North America Summer 2006 • North America “Problem” (OD/RD special topic 2005) • strong, moist warm flow from the Gulf of Mexico • large and divergentwind increments at 150-250 hPa • the conclusion was that “increments are not related to bad observations or a poor 4D-Var performance”

14. Summer case 2006 ERA40 Jan ERA40 Jun Mean 850-hPa Wind & Z500 hPa courtesy by Fernando Prates Mean CAPE Mean TCWV

15. FSO showed a Fc Error increase due to the American wind profiler observations. Southerly flow across SE USA bringing warm and moist air from Gulf of Mexico produced strong convective instability in the region, a typical situation at this time of the year. Following Ackley et al report (1998) on wind profiler measurements validity “in strong unstable conditions (turbulence) the measure of the mean horizontal wind is corrupted affecting the measurements”. Suggesting that the forecast impact can change with the meteorological situation for the summer 2006 case. Summary FSO wind Profiler

16. FSO: Atmospheric Motion Vector FcE Contribution Summer 2006 Forecast error contribution of the observed wind grouped by satellite types- positive corresponds to an increase of Fc Error Forecast error contribution of the wind on pressures levels & grouped by satellite types- largest degradation comes from the lower troposphere

17. FSO AMV 700-1000 hPa U-Wind: Summer 2006 RMSE AMV- Baseline 850 hPa U

18. FSO AMV 700-1000 hPa Summer 2006 U-comp V-comp Mean 850hPa wind Atlantic Ocean: transition between sub-tropical and extra-tropical from weak to strong zonal flow Indian Ocean: well established Monsoon circulation courtesy by Fernando Prates

19. AN mean vertical velocity (*0.01 Pa/s) ERA40 south south north north FSO Atlantic Ocean: Observation Quality Cross Section [35W-0E] The strong sinking motion in SH near 30S represents the southern limit of the Hadley circulation where the subtropical high cell is located. Cloud suppression or low clouds. AMV quality: difficult to assign the height of the cloud top courtesy by Fernando Prates

20. FSO Indian Monsoon Summer 2006: Model bias u-wind v-wind A too strong low level flow of Indian Summer Monsoon is a well known problem in the model as is indicated by the JJA mean analysis increments Diagnostic explorer Mean An inc 925-hPa JJA 2006

21. Positive impact negative impact AMV FSO 700-1000 hPa: Winter 2007 Overall impact of the observations to FcError Largest negative impact of AMVs to Fc error can be seen in central/eastern Pacific (absent in summer case). Negative impact seen during summer 06 in south Atlantic near 30S has disappeared in winter 07 In the Indian Ocean the degradation is mainly due to u-component of the wind v-wind u-wind

22. ERA40 Mean vertical velocity (*0.01 Pa/s) 180-150 W north south • A second cluster of negative impact near 25N/140W is localized on top of a region of weak winds (strong sinking motion/ high pressure system) 180W 150W 135W u-wind Winter 2007 central/eastern PacificCross Section • The largest negative impact of AMVs to the Fc error is found between 5N - 15N coinciding to a broad downward mean motion of the Hadley circulation. Large departures were also found below 700hPa in the same region.

23. FSO showed a Fc error increase due to AMVs The location of the largest negative impact of the AMVs in Atlantic (Summer 2006) and in pacific (Winter 2007, El Nino) is found close to the region of strong sinking mean motion embedded in the Hadley circulation Observation quality problem on the height assignment Detrimental effect is also observed in the Indian ocean associated with a too strong Indian monsoon circulation developed by the model Model bias Summary FSO AMVs

24. 48h Fce 24h Fce GPS RO Impact on Forecast Error Winter 2007

25. 6000 Automatic&Manual Surface Press SYNOP FcE Conributionwinter case SYNOP sfc-press observations shows an overall globally positive impact to the forecast error but not over Europe.

26. Eu Area manual Eu Area metar Daily Fc error contribution over Europe Automatic Surf Press SYNOP FcE Contribution time series - Winter 2007Storm Kyrill – 18 -20 Jan

27. Overall impact of the observations to fc error hPa 10 40 500 GPS RO Impact on Forecast Error Winter 2007 GPS RO at 50-hPa The negative impact is more pronounced in the tropics & subtropics

28. RMSE Mean 50 hPa RMSE Temperature GPSRO-Control Winter 2007 50-hPa Temp RMSE differences between GPS RO-Control OSEs (24-hrs Fc) The degradation (positive values) are found mainly in the tropical belt which is consistent with the geographical distribution obtained from the FSO GPS-RO Control The OSE shows a positive impact for the GPS-RO for the 10-days forecast with the exception of the first 24hrs forecast.

29. A negative impact to Fc error due to GPS-RO is found in the lower stratosphere and mainly in the tropical belt which is related with temperature model bias. OSE showed the same impact for the first 24hrs forecast but also the positive impact for longer time ranges. The overall decrease of Fc error due to SYNOP (man. & auto.) contrasted with the degradation over Europe. Adverse weather conditions over Europe (strong pressure gradient) for several weeks would require a higher resolution analysis system. Summary FSO GPS-RO and SYNOP/METAR sfc-pressure

30. xb y y xb xb y Analysis SolutionModel space Observation space B(qxq)=Var(xb) R(pxp)=Var(y) K(qxp) gain matrix H(pxq) Jacobian matrix Analysis Sensitivity to the Observation or Information Content FSO Cardinali 2009 Cardinali et al 2004

31. Operational ECMWF system September to December 2008 All Observations

32. Over the last decade the assessment of each observation contribution to analysis and forecast is among the most challenging diagnostics in data assimilation and NWP. Forecast sensitivity to observations allows to monitor the observation forecast impact on the 24 range The tool provides information on the observation type, subtype, variable and level responsible for the forecast error variation. Causes must be found that explain the failure Failures can be due to the data quality or some characteristics of the assimilation system and can highly depend on the weather situation A joint effort blending different expertises, tool developers and meteorologists, is necessary to produce a comprehensive investigation and understanding of forecast failures The assessment should be carried out on a daily basis Conclusion&Remarks