Impact of increased satellite data density in sensitive areas

1. Impact of increased satellite data density in sensitive areas Carla Cardinali, Peter Bauer, Roberto Buizza, Jean-Noël Thépaut European Centre for Medium-Range Weather Forecasts in lovely Reading, Berkshire, UK 3rd THORPEX International Science Symposium 09/2009

2. Study contents • Background • Thinning of data is applied to: • reduce data volume • avoid the introduction of spatial observation error correlation that is currently not accounted for in data assimilation algorithm • Thinning is performed statically on a fixed latitude/longitude grid. • Objective • Evaluate impact of selective satellite observational data thinning on medium-range NWP aiming at denser data in sensitive areas and less dense data in other areas  trade-off between data impact and data volume. • Approach • Experiments with global data thinning: • Change global latitude/longitude thinning grid. • Experiments with data thinning in selected regions: • Increased density in sensitive areas and reduced density elsewhere using a Singular Vector based measure to identify areas from which forecast errors are growing fast (ECMWF 2007 QJ papers). • Sensitive areas are computed for Southern Hemisphere 3rd THORPEX International Science Symposium 09/2009

3. Experiments • Global data thinning • Thinn_2.5 : global latitude/longitude thinning is 2.5o. • Thinn_1.25 : ~ is 1.25o current operational configuration • Thinn_0.625 : ~ is 0.625o. • Selective data thinning • Thinn_Cntrl : ~ is 1.25oproxy for Thinn_1.25 • Thinn_SV : ~ is 1.25o and 0.625o in SV areas. • Thinn_RD : ~ is 1.25o and 0.625o in randomly distributed areas. • Thinn_CSV : ~ is 1.25o and 0.625o in Climatological SV areas. • Thinn_0.625 • Additional information • All experiments are run at T511L91 (12-hour 4D-Var) for 01/12/2008-28/02/2009. • All experiments are verified with T799L91 operational model analyses (without first 7 days (spin-up) i.e. 83 cases). • All SV/RD/CSV areas occupy same fraction (15%) of the Southern hemisphere. • The SV-based climatology derived from the mean 2007 SV-areas. 3rd THORPEX International Science Symposium 09/2009

4. Global data thinning: Forecast impact Thin_2.5-Thin_1.2 500hPa Normalized RMSE 95% confidence 55 cases North H South H 0 1 2 3 4 5 6 7 8 Forecast Day 3rd THORPEX International Science Symposium 09/2009

5. 0 1 2 3 4 5 6 7 8 Forecast Day Global data thinning: Forecast impact Thin_1.2-Thin_0.6 500hPa Normalized RMSE 95% confidence 83 cases North H South H 3rd THORPEX International Science Symposium 09/2009

6. Data coverage: Single case 01/01/2009 00 UTC data density AMSU-A channel 9: Singular Vectors: Randomly distributed circular areas: 2007 Singular Vector climatology: 3rd THORPEX International Science Symposium 09/2009

7. Data coverage: Average 01-07/01/2009 00 and 12 UTC data density AMSU-A channel 9: Singular Vectors: Randomly distributed circular areas: 2007 Singular Vector climatology: 3rd THORPEX International Science Symposium 09/2009

8. Selective data thinning: DFS Decrease of DFS relative to the Thin_0.625 experiment Global Southern Hemisphere 3rd THORPEX International Science Symposium 09/2009

9. Selective data thinning: Forecast impact SV-CNTRL Southern H. Normalized RMSE 95% confidence 83 cases 1000 hPa 500 hPa 200 hPa 0 1 2 3 4 5 6 7 8 Forecast Day 3rd THORPEX International Science Symposium 09/2009

10. Selective data thinning: Forecast impact SV-RD Southern H. Normalized RMSE 95% confidence 83 cases 1000 hPa 500 hPa 200 hPa 0 1 2 3 4 5 6 7 8 Forecast Day 3rd THORPEX International Science Symposium 09/2009

11. Selective data thinning: Forecast impact SV - CSV Southern H. Normalized RMSE 95% confidence 83 cases 1000 hPa 500 hPa 200 hPa 0 1 2 3 4 5 6 7 8 Forecast Day 3rd THORPEX International Science Symposium 09/2009

12. Conclusions • Global data thinning • Current operational setting (1.25o) is too conservative and satellite data density could be increased. • 0.625 º thinning will be considered for operational implementation. • Selective data thinning • Forecast scores are best for experiment with increased data density in SV-based areas that are updated for each analysis. • 40% loss of DFS by increasing the data density over SV areas instead than globally. • 0.42° thinning is thought to be used in SV areas. 3rd THORPEX International Science Symposium 09/2009