2. TransCom 3 Level 2 Base Case Inter-annual CO 2 Flux Inversion Results Current Status David Baker, Rachel Law, Kevin Gurney, Peter Rayner, TransCom3 L2 modelers*, and the producers of the GLOBALVIEW-CO 2 data product * (P. Bousquet, L. Bruhwiler, Y-H Chen, P. Ciais, I. Fung, K. Gurney, M. Heimann, J. John, T. Maki, S. Maksyutov, P. Peylin, M. Prather, B. Pak, S. Taguchi, Z. Zhu) 14 June 2004

3. Outline TransCom3 and the Level 2 inter-annual inversion Inter-annual inversion results –Changes in base case since Nov 2002 –Inter-annual variability (IAV) results Significance: are transport and inversion uncertainties small enough to see real IAV? Robustness to set-up: sensitivity to the tightness of the ocean prior and to the SEY data uncertainty Mean and seasonal flux results

4. TransCom3 Background In past decade: large spread in regional CO 2 flux estimates from published inversions! Differences in the problem set-up (time span, measurement stations, data uncertainties, a priori fluxes and flux uncertainties, etc.) cause some of this spread. BUT, it was thought that differences between transport models might be responsible for the largest differences. Goal of TransCom3: find the sensitivity of CO 2 flux inversion results to the transport model used. [Assumption: no interannual variability in winds; each model uses only one year of “typical” winds.] ALSO: are there flux results that are robust across all (or most) of the transport models?

5. TransCom 3 Three types of inversions have been done as part of T3 levels 1 & 2. All use Bayesian synthesis, but make different assumptions about the time history of the fluxes and use data averaged over different spans. Inversion TypeFluxesDataReferences Long-term meanRSGurney, et al, Nature 415, 626-630 SeasonalR*12S*12Gurney, et al. (2004) Inter-annualR*12*YS*12*Y Rayner, et al, Tellus, 51B, 213-232; Bousquet, et al., Science, 290, 1342-46; R = 22 regions S = 78 stations (GLOBALVIEW-CO 2, 2003) Y = 15 years (1988-2002) solved for here

6. Base Case Assumptions Nov 2002 [for T3 L3] 1988-2001 (14 years) GLOBALVIEW- CO 2 (2002), 76 sites [chosen to have >68% data coverage ]; interpolated data used to fill all gaps. Data uncertainties calculated from GV 1979-2002 rsd (  GV ) as:  2 = (0.3 ppmv ) 2 +  GV 2 [non-seasonal] June 2004 1988-2002 (15 years) GLOBALVIEW- CO 2 (2003), 78 sites; the previous 76 + CPT_36C0 + HAT_20C0; also SYO_00D0 changed to SYO_09C0 New seasonally- and interannual-varying data uncertainties

7. Base Case Assumptions Nov 2002 [for T3 L3] A priori fluxes – same as in Level 1, constant across year A priori flux errors – twice Level 1 June 2004 A priori fluxes – Kevin’s seasonally-varying ones from the seasonal inversion A priori flux errors a) Kevin’s seasonally- varying ones b) ditto for land regions,  2 =  2 L1 + (0.5 PgC/yr ) 2 for ocean regions

9. Method Find optimal fluxes x to minimize where: x are the CO 2 fluxes to be solved for, H is the transport matrix, relating fluxes to concentrations z are the observed concentrations, minus the effect of pre-subtracted tracers (fossil fuel, and seasonal CASA & Takahashi) R is the covariance matrix for z, x o is an a priori estimate of the fluxes, P xo is the covariance matrix for x o Solution:

10. Time-dependent basis functions for 13 transport models were submitted in Level 2: CSU (Gurney) † GCTM (Baker) GISS-UCB (Fung) GISS-UCI (Prather) JMA-CDT (Maki) MATCH (Chen) † not used here 12 + 1 – 1 = 12 models used MATCH (Law) MATCH (Bruhwiler) NIES (Maksyutov) NIRE (Taguchi) TM2 (LSCE) TM3 (Heimann) PCTM (Zhu)

11. Inter-annual Variability (IAV) Results: Key Issues Is the transport error low enough that key features in the inter-annual variability can be robustly identified by region? Is the random estimation error low enough? Is the IAV robust to the basis functions used? [Frequency- and SVD-truncation used to test this in past; here we examine the effect of a tighter ocean prior, and adjusting the data error for SEY] Where is the variability strongest and most robust? What physical mechanisms might cause it?

12. EUROPE: Monthly Flux EUROPE: Deseasonalized Flux 12-model median flux

13. EUROPE: Deseasonalized Flux EUROPE: Deseasonalized Flux, Mean Subtracted Off 1-sigma model spread 1-sigma internal error 12-model median flux (Summary plot)

14. Computation of the inter-annual variability (IAV), long-term mean, and seasonality from the monthly estimate, x mon x mon = x deseas + x seas = x mean + x IAV + x seas x deseas computed by passing a 13-point running mean over x mon x seas = x mon - x deseas (zero annual mean seasonal cycle) x mean = the 1988-2002 mean of x deseas x IAV = x deseas - x mean (zero mean, 1988-2002) Corresponding errors also computed

15. Chi-square Significance Test We try to reject the null hypothesis that the estimated IAV is due solely to the combined effect of both transport error and random estimation error, superimposed on zero IAV Compare the variance of x IAV with the combined variance the transport and random errors: use  2 test ( =14; 15 independent years – 1 for mean)

16. 4 cases presented here, to illustrate the impact of regularization (by tightening the ocean prior), and sensitivity to SEY: Loose ocean: Kevin’s a priori ocean errors Tight ocean:  2 =   Level 1 + (0.5 PgC/yr ) 2 Tight ocean, loose SEY: add 1.5 ppmv in quadrature to Seychelles data error, 1988-96 Tight ocean, loose SEY*: ditto, with an error in the a priori uncertainties for May-Nov for Region 3 corrected (increased to Kevin’s values)

17. Total Flux (Land+Ocean) Loose ocean errors <0.00001

18. Total Flux (Land+Ocean) Tight ocean errors <0.00001

19. Total Flux (Land+Ocean) Tight ocean, loose SEY <0.00001

20. Total Flux (Land+Ocean) Tight ocean, loose SEY * <0.00001

21. Land & Ocean Fluxes Loose ocean errors <0.00001 0.00001 <0.00001 (0.85)

22. Land & Ocean Fluxes Tight ocean errors <0.00001 0.00004 <0.00001 0.0034(0.31)

23. Land & Ocean Fluxes Tight ocean, loose SEY (0.38) <0.00001 0.000036 0.00007 0.0062

24. Land & Ocean Fluxes Tight ocean, loose SEY * <0.00001 0.000036 0.00013 0.0057(0.37)

25. Loose ocean errors 0.0021 0.09 0.019 0.069

26. Tight ocean errors 0.00001 0.055 0.006 (0.114)

27. Tight ocean, loose SEY 0.00002 0.051 0.004 (0.115)

28. Tight ocean, loose SEY * 0.00003 0.052 0.006 (0.116)

29. Loose ocean errors <0.00001 0.0042 0.053 0.025

30. Tight ocean errors <0.00001 0.00016 <0.00001 0.0055 (0.13)

31. Tight ocean, loose SEY <0.00001 0.00015 <0.00001 0.0016

32. Tight ocean, loose SEY * <0.00001 0.00014 <0.00001 0.0022

33. Loose ocean errors 0.097 (0.41) <0.00001

34. Tight ocean errors (0.25) (0.155) <0.00001

35. Tight ocean, loose SEY (0.24) (0.137) <0.00001

36. Tight ocean, loose SEY * (0.23) (0.156) <0.00001

37. Loose ocean errors

38. Tight ocean errors

39. Tight ocean, loose SEY

40. Tight ocean, loose SEY *

41. “But wait – you aren’t allowed to change the data errors on SEY just because you don’t like the estimate you get for the Tropical Indian!!” True, in general, unless you have good reason to believe you used overly tight data errors for the site before… Tom Conway, NOAA CMDL (pers. comm., 5/11/04): –“Seychelles started out pretty good, but then we had various problems over the years. I think the most recent 8 years are pretty good again.” –“Things look pretty bad again in 1989 and 1990…. `91, `92, and `93 look pretty good…” –“In 1994… the USAF took over. The samples, most of which I believe had been collected near the coast, now were collected by USAF personnel at the tracking station (inland). This is the period when the sample collectors wrote down 270° for the wind direction of almost every sample.” [The USAF pulled out in 1996, and a measurer trained by CMDL began taking samples.]

42. Comparison of our 1992-96 Mean Fluxes (right) to Level 1 (left)

43. Mean Seasonal Cycle 1991-2000 Prior Prior, no def. G04 1992-96

44. Mean Seasonal Cycle 1991-2000 Prior G04 1992-96

45. Seasonal Cycle Amplitude [PgC/yr]

46. Conclusions Inter-model differences in long-term mean fluxes are larger than in the flux inter-annual variability IAV for latitudinal land & ocean partition is robust (except for Southern S. America); continent/basin partition of IAV in north is of marginal significance; in tropics, IAV is significant for the Tropical Pacific and Australasia The IAV for the 22 regions is significant for only a few land regions and about half the ocean regions. Probable physical drivers for Tropical Asia (fires) & East Pacific (El Niño); other regions less clear… Good agreement between the three types of inversions (annual-mean, seasonal, inter-annual) in mean & seasonality