1. - 2- 2 The 2 nd phase of the Global Land-Atmosphere Coupling Experiment Randal Koster GMAO, NASA/GSFC randal.d.koster@nasa.gov

2. GLACE-1 was a successful international modeling project that looked at soil moisture impacts on precipitation... Where precipitation responds to variations in soil moisture, according to the 12 GLACE models.

3. Motivation for GLACE-2 For soil moisture initialization to add to subseasonal or seasonal forecast skill, two criteria must be satisfied: 1.An initialized anomaly must be “remembered” into the forecast period, and 2.The atmosphere must be able to respond to the remembered anomaly. Addressed by GLACE Addressed by GLACE2: the full initialization forecast problem

4. GLACE-2 : Experiment Overview Perform ensembles of retrospective seasonal forecasts Initialize land states with “observations”, using GSWP approach Prescribed, observed SSTs or the use of a coupled ocean model Initialize atmosphere with “observations”, via reanalysis Evaluate forecasts against observations Step 1:

5. GLACE-2 : Experiment Overview Perform ensembles of retrospective seasonal forecasts Initialize land states with “observations”, using GSWP approach Prescribed, observed SSTs or the use of a coupled ocean model Initialize atmosphere with “observations”, via reanalysis Evaluate forecasts against observations Step 2: “Randomize” land initialization!

6. GLACE-2 : Experiment Overview Step 3: Compare skill; isolate contribution of realistic land initialization. Forecast skill obtain in experiment using realistic land initialization Forecast skill obtained in identical experiment, except that land is not initialized to realistic values Forecast skill due to land initialization

7. FORECAST START DATES Apr 1 1986 1987 1988 1989 1990 100 different 10-member forecast ensembles. Each ensemble consists of 10 simulations, each running for 2 months. 1991 1992 1993 1994 1995 Apr 15 May 1 May 15 Jun 1 Jun 15 Jul 1 Jul 15 Aug 1 Aug 15

8. Observed precipitation Wind speed, humidity, air temperature, etc. from reanalysis Observed radiation LSM A decade of LSM initial conditions for seasonal forecasts The resulting LSM initial conditions reflect observed antecedent atmospheric forcing. LAND STATE INITIALIZATION via offline simulations, a la GSWP A decade of offline integration

9. Two goals of project: 1. Calculate “potential predictability” in each forecast system – where does atmospheric noise overwhelm any potential signal? 2. Calculate skill in P and T prediction associated with the accurate initialization of land surface states.

10. Progress to date…

11. Integration into WCRP/TFSP project (TFSP = Task Force on Seasonal Prediction) {

12. Updated Participant List Group/ModelPoints of Contact 1. NASA/GSFC (USA): GMAO seasonal forecast system (old and new) 2. COLA (USA): COLA GCM, NCAR/CAM GCM 3. Princeton (USA): NCEP GCM 4. IACS (Switzerland): ECHAM GCM 5. KNMI (Netherlands): ECMWF 6. ECMWF 7. GFDL (USA): GFDL system 8. U. Gothenburg (Sweden): NCAR 9. CCSR/NIES/FRCGC (Japan): CCSR GCM 10. FSU/COAPS # models S. Seneviratne, A. Roesch E. Wood, L. Luo P. Dirmeyer, Z. Guo R. Koster, T. Yamada2 B. van den Hurk T. Gordon J.-H. Jeong T. Yamada 2 1 1 1 1 1 1 12 models 1 G. Balsamo M. Boisserie1

13. “Persisted” SST boundary conditions have been constructed and are now available online. (T. Yamada)

14. Fcst. ModelPoints of ContactProgress to Date COLA GCM ; NCAR/ CAM GCM, via COLA Paul Dirmeyer, Zhichang Guo -- Forcing data interpolated to proper resolution; offline land simulations proceeding. -- Completed 10 years of COLA runs -- NCAR runs being set up. Three analyses performed here at GLACE-2 Central on COLA results: 1. Analysis of potential predictability. 2. Analysis of forecast skill. 3. Analysis of forecast skill, using statistical enhancements.

15. For a given ensemble forecast, assume that the first ensemble member represents “nature”. STEP 1: Assume that the remaining ensemble members represent the “forecast”. STEP 2: STEP 3: Determine the degree to which the “forecast” agrees with the assumed “nature”. Precipitation anomaly (mm/day) 0 1 2 3 4 -2 -3 -4 Ensemble member 1 23 4 56789 To what extent does this anomaly… … agree with the average of these anomalies? 10 Potential predictability is the maximum predictability possible in the forecasting system.

16. Potential Predictability (r 2 ), Precipitation (COLA model, Series 1) (July 1 start, land initialized) 1 st half of July2 nd half of July 1 st half of August2 nd half of August

17. Potential Predictability (r 2 ), Precipitation (COLA model, Series 2) (July 1 start, land initialized) 1 st half of July2 nd half of July 1 st half of August2 nd half of August

18. Net land impact on Potential Predictability (r 2 ), Precipitation (COLA model, Ser1 – Ser2) (July 1 start, land initialized) 1 st half of July2 nd half of July 1 st half of August2 nd half of August

19. Forecast Skill: COLA model Notes: Skill is averaged separately for four leads: a. 1-15 days after forecast start date b. 16-30 days after forecast start date c. 31-45 days after forecast start date d. 46-60 days after forecast start date For a given forecast start date and lead, the forecasts from the 10 ensemble members are averaged into a single field. Skill for a given period is compared to observations during that period. (Measured as r 2.) Analysis focuses on US first, out of convenience: we have access to a high quality multi-decade observational dataset there (Higgins et al., 2000), and besides, most models show some coupling strength there (GLACE-1). Prior to computing the skill scores (observations), all 15-day forecasts are standardized (using relevant means and standard deviations for given start date and lead), as are all the observations.

20. Forecast Skill: COLA model Notes (continued): The skill analysis is supplemented with an analysis of transformed forecasts: Suppose we can predict P here with the forecast system (high potential predictability)… Suppose, though, that P in these two areas are correlated in the real world. Then we can combine the model forecasts with the observed correlations to derive a forecast here. …but not here Using these ideas, we can compute a “transformation matrix” A that improves a forecast: x = A x ~ Vector holding original forecasts Vector holding transformed forecasts For details, see Koster et al., Monthly Weather Review, 136, p. 1923-1939.

21. Forecast skill (COLA): Precipitation All start dates (100 standardized values going into r 2 calculation). Lead: Days 1-15 Series 2: no land information Series 1: land information Differences: impact of land Raw Results Transformed Results (colors go from 0. to 0.5)(colors go from -0.5 to 0.5)

22. Forecast skill (COLA): Precipitation All start dates (100 standardized values going into r 2 calculation). Lead: Days 15-30 Series 2: no land information Series 1: land information Differences: impact of land Raw Results Transformed Results (colors go from 0. to 0.5)(colors go from -0.5 to 0.5)

23. Forecast skill (COLA): Precipitation AMJ start dates (60 standardized values going into r 2 calculation). Lead: Days 1-15 Series 2: no land information Series 1: land information Differences: impact of land Raw Results Transformed Results (colors go from 0. to 0.5)(colors go from -0.5 to 0.5)

24. Forecast skill (COLA): Temperature Raw results, all start dates Series 2: no land information Series 1: land information Differences: impact of land Lead: 1-15 days (colors go from 0. to 0.5) (colors go from -0.5 to 0.5) Lead: 16-30 days Lead: 31-45 days Lead: 46-60 days

25. Fcst. ModelPoints of ContactProgress to Date NCAR (USA, via U. Gothenburg, Sweden) Jee-Hoon Jeong-- Baseline set of simulations for the period 1986-1995 is finished (Series 1 and Series 2). -- Performing additional forecasts with modified initialization strategy. We found some unusual aspects of the results that need clearing up – we need to talk to Jee-Hoon. Currently, the results for NCAR are indeterminate.

26. Fcst ModelPoints of ContactProgress to Date GEOS5 GCM; NSIPP GCM (NASA/GSFC) Randal Koster, Tomohito Yamada -- Simulated 50 years of land surface conditions for initialization -- Ran GEOS5 GCM 10 years to generate climatology -- July 1 forecasts finished.

27. Model: GEOS5 Variable: SWOI Start date: July 1 (potential predictability) Ser1 Ser2 Ser1-2

28. Model: GEOS5 Variable: PRCP Start date: July 1 (potential predictability) Ser1 Ser2 Ser1-2

29. Model: GEOS5 Variable: PRCP Start date: July 1 (potential predictability) Ser1 Ser2 Ser1-2 These runs used an early, unfinished version of the NASA free-running climate model, a version known to be deficient in land-atmosphere coupling strength.

30. Fcst. ModelPoints of ContactProgress to Date KNMI Bart van den Hurk, Helio Camargo, Gianpaolo Balsamo -- GSWP2 forcings regridded to their GCM’s resolution. -- 10-yr climatology run with the GCM, to allow for soil moisture scaling. -- Land model incorporated into LIS, for efficient offline simulation. -- 1+ years of Series 1 forecasts, 1 set of Series 2 forecasts. (Forecasts are ongoing.) First results showing forecasted soil moisture’s agreement with “truth” across the globe: -- decrease of agreement with time -- agreement differs amongst ensemble members. -- longer apparent memory in mid-summer

31. Fcst ModelPoints of ContactProgress to Date GFDL (USA) Tony Gordon-- AMIP style control run performed for atmospheric initial conditions and for scaling of land variables. -- 10 years (1 st of each of month, 10 ensemble members) completed, for both Series 1 and Series 2. -- All Series 1 runs done; scaled and unscaled; Series 2 done two ways: with pdf, and with average.

32. NCEP (via Princeton, USA) Eric Wood, Lifeng Luo -- Simulated 50 years of land surface conditions for initialization. -- Ready to go; waiting for time on NCEP machine. ECHAM (via IACS, Switzerland) Sonia Seneviratne, Roesch Andreas -- Series 2 simulations for GSWP2 period are finished for most start dates in 10-year period. ECMWF Gianpaulo Balsamo CCSR/NIES/ FRCGC (Japan) Tomohito Yamada-- Simulated 50 years of land surface conditions for initialization. FSU/COAPS Marie Boisserie(New to project) Fcst ModelPoints of ContactProgress to Date

33. Discovered: Problem with GLACE2 SST data (Thanks to Tony Gordon for spotting this.) Original data: Hadley monthly mean state Two types of data 1. start date: 1 st (interpolated by two months, e.g., (March+April)/2=April 1st) 2. start date: 15 th (directly from Hadley monthly mean sate at a same month)

34. 1995, Sea Surface Temperature GLACE2 (black line) & Hadley (Green line) 15 th simulation1st simulation GLACE2: April 1 st Hadley: (March+April)/2 GLACE2: April 1 st Hadley: (April+May)/2 Main point: Somehow, the SST data meant to refer to March 1 actually refers to observations on April 1. The problem only applies to start dates on the first of the month. SST data for start dates on the 15 of the month are not in error. New SST datasets are being constructed now.

35. 1995, Sea Surface Temperature GLACE2 (black line) & Hadley (Green line) 15 th simulation1st simulation GLACE2: April 1 st Hadley: (March+April)/2 GLACE2: April 1 st Hadley: (April+May)/2 Main point: Somehow, the SST data meant to refer to March 1 actually refers to observations on April 1. The problem only applies to start dates on the first of the month. SST data for start dates on the 15 of the month are not in error. New SST datasets are being constructed now. Problem has been fixed!

36. Summary: Aside from that glitch, GLACE-2 is proceeding nicely!

37. Back-up slides

38. Computational requirement: 100 forecast start dates 10 ensemble members per forecast 1/6 years (2 months) per simulation 2 experiments (including control) XXX = 333 years of simulation

39. Required output diagnostics (to be provided to GLACE data center): 1. For each of the 4 15-day periods of each forecast simulation, provide global fields of: 15-day total precipitation 15-day average near-surface air temperature (in the lowest AGCM level) 15-day total evaporation 15-day average net radiation 15-day average vertically-integrated soil moisture content. 15-day average near-surface relative humidity

40. Required output diagnostics (to be provided to GLACE data center): 2. For each day of the first 30 days of certain forecast simulations, provide global fields of: vertically-integrated soil moisture content. This will allow us to analyze the decay with lead time of the information provided by soil moisture initialization. The forecast simulations chosen for this output are: Apr. 1, May 1, June 1, July 1, and Aug. 1 of 1986, 1988, 1990, 1992, and 1994.

41. 3. Generation of SST Boundary Conditions Needed: SST conditions for forecasts that do not include measurements of SSTs during the forecast period. Approach: Determine persistence timescales from observational record (with data exclusion) and reduce initial (measured) SST anomalies with time into the forecast, using those timescales.

42. “Persisted” SST boundary conditions have been constructed and are now available online. (T. Yamada)

43. GSWP2 Princeton June 1988 Soil Moisture Anomaly (against 1986-1995) 4. Examined forcing datasets (T. Yamada)

44. 1. Determine maximum level of precipitation (and air temperature) predictability associated with land initialization. Results from pilot study of Koster et al. 2004

45. Regress full forecast against actual observations to retrieve r 2. 2. Determine forecast skill for precipitation (and air temperature) associated with land initialization.

46. Contribution to skill: P forecasts Contributions of land moisture initialization to the skill of subseasonal (one-month) forecasts of P and T Contribution to skill: T forecasts Koster et al., Journal of Hydrometeorology,5, pp. 1049-1063, 2004 Results from pilot study of Koster et al. 2004