Verification of a downscaling approach for large area flood prediction over the Ohio River Basin N. Voisin, J.C. Schaake and D.P. Lettenmaier University of Washington, Seattle, WA AMS Annual Meeting, Phoenix AZ 11-15 Jan 2009
Objective Predict streamflow and associated hydrologic variables, soil moisture, runoff, evaporation and snow water equivalent : Applicable to large river basins, eventually globally: spatial consistency, ungauged basins Using a fully distributed hydrology model Using ensemble weather forecasts Lead time up to 2 weeks
Objective Forecast schematic BCSD = Bias correction and statistical downscaling Forecast schematic Several years back Medium range forecasts (2 weeks) ECMWF EPS 50 ensemble members 2002-2008 Daily ERA-40 surrogate for near real time analysis fields 1979-2002 Daily ECMWF Analysis 2002-2008 BCSD to 0.25 degree BCSD with forecast calibration, 0.25 degree Atmospheric inputs VIC Hydrology Model Hydrologic model spinup 0.25 degree Hydrologic fcst (stream flow, soil moist., SWE, runoff ) Initial State Flow fcst calibration
Objective Compare different downscaling techniques Applicable at a global scale For precipitation forecast Improve or conserve the skill
Outline Existing downscaling methods Analog technique and various variations of it Forecast Verification at different spatial and temporal scales: Mean errors Predictability, reliability Spatial rank structure
1. Downscaling techniques MOS (Glahn and Lowry 1972, Clark and Hay 2004) Bias correction followed by spatial and temporal resampling for seasonal forecast (Wood et al. 2002 and 2004) National Weather Service (NWS) Ensemble Precipitation Processor (EPP) ( Schaake et al. 2007) Analog techniques ( Hamill and Whitaker 2006)
2. Analog technique 3 methods for choosing the analog: ( adapted from Hamill and Whitaker 2006) Retrosp. FCST dataset, +/- 45 days around day n 1 degree resolution Corresp. Observation (TRMM) 0.25 degree resolution FCST D DAY OBS D DAY Downscaled FCST day n 0.25 degree FCST D DAY OBS D DAY FCST D DAY OBS D DAY FCST D DAY OBS D DAY FCST D DAY OBS D DAY FCST D DAY OBS D DAY FCST D DAY OBS D DAY FCST D DAY OBS D DAY FCST D DAY OBS D DAY FCST D DAY OBS D DAY FCST D DAY OBS D DAY FCST D DAY OBS D DAY FCST n +/- 45 days Year-1 OBS n +/- 45 days Year-1 FCST day n 1 degree 5 degree 3 methods for choosing the analog: Closest in terms of RMSD, for each ensemble 15 closest in terms of RMSD, to the ensemble mean fcst Closest in terms of rank, for each ensemble 5 degree
2. Analog technique Spatial domain for the analog Choose an analog for the entire domain (Maurer et al. 2008): entire US, or the globe Ensure spatial rank structure Need a long dataset of retrofcst-observation. Moving spatial window (Hamill and Whitaker 2006): 5x5 degree window (25 grid points) Choose analog based on ΣRMSD, or Σ(Δrank) Date of analog is assigned to the center grid point
2. Analog technique Ens. Mean Fcst, 20050713 Fcst 20050713 4 closest analogs in the retrospective forecast dataset Corresponding 0.25 degree TRMM for the analogs, Downscaled ensemble forecast members Downscaled ens. mean forecast TRMM (obs) ( adapted from Hamill and Whitaker 2006)
3. Forecast Verification Evaluate the different analog techniques, simple interpolation, and basic resampling downscaling Verification conditioned on the forecast: Mean errors Reliability Predictability Verification conditioned on the observation Discrimination (ROC) For lead times 1,5 and 10 days at 0.25 and 1 degree spatial resolution, Daily and 5 day accumulation
Mean Errors 0.25 degree Ohio Basin 2002-2006 TRMM as obs Upper tercile: improved bias
Reliability of ens. spread 0.25 degree Ohio Basin 2002-2006 TRMM as obs Improved reliability
Predictability 0.25 degree Ohio Basin 2002-2006 TRMM as obs Status quo or no improvement
Discrimination ROC diagram 0.25 degree Ohio Basin 2002-2006 TRMM as obs Prob. of detection Or hit rate False alarm rate
Spatial structure 2005, Jul 13th 75th Percentile basin daily acc., 2002-2006 TRMM
Conclusions The analog technique with a moving spatial window improves: reliability (considerably), mean errors (slightly) Status quo on: discrimination,predictability Results consistent at different spatial and temporal scales ( not shown, 1 degree and 5 day acc.) More realistic precipitation patterns. Spatial rank structure? An analog technique with no moving spatial window would ensure it. Issue with short observed dataset. Try the NWS EPP.
Climatologies of forecasts Ohio Basin 2002-2006
Mean Errors 0.25 degree Ohio Basin 2002-2006 TRMM as obs Upper tercile: improved bias
Mean Errors 1 degree Ohio Basin 2002-2006 TRMM as obs Upper tercile: improved bias
Mean Errors 0.25 degree 5 day acc. Ohio Basin 2002-2006 TRMM as obs Upper tercile: improved bias
Reliability 0.25 degree Ohio Basin 2002-2006 TRMM as obs - Improved reliability - poor reliability for medium tercile - poor reliability lead time 10
Reliability 1 degree Ohio Basin 2002-2006 TRMM as obs - Improved reliability - No reliability for medium tercile - No reliability lead time 10
Reliability 0.25 degree 5 day acc Ohio Basin 2002-2006 TRMM as obs - Improved reliability No reliability for medium tercile - Some reliability day 6-10
Sharpness 0.25 degree Ohio Basin 2002-2006 TRMM as obs Improved sharpness for lower tercile
Sharpness 1 degree Ohio Basin 2002-2006 TRMM as obs Improved sharpness for lower tercile
Sharpness 0.25 degree 5 day acc Ohio Basin 2002-2006 TRMM as obs No improvement
Predictability 0.25 degree Ohio Basin 2002-2006 TRMM as obs Status quo or no improvement
Predictability 1 degree Ohio Basin 2002-2006 TRMM as obs Status quo or no improvement
Predictability 0.25 degree 5 day acc Ohio Basin 2002-2006 TRMM as obs Status quo or no improvement
Reliability of ens. spread 0.25 degree Ohio Basin 2002-2006 TRMM as obs
Reliability of ens. spread 1 degree Ohio Basin 2002-2006 TRMM as obs
Reliability of ens. spread 0.25 degree 5 day acc. Ohio Basin 2002-2006 TRMM as obs