Spatial Processes and Land-atmosphere Flux

1. Spatial Processes andLand-atmosphere Flux Constraining ecosystem models with regional flux tower data assimilation Flux Measurements and Advanced Modeling, 22 July 2008 CU Mountain Research Station, “Ned”, Colorado Ankur Desai Atmospheric & Oceanic Sciences, University of Wisconsin-Madison

2. Let’s get spacey…

3. And regional

4. Why regional? • Spatial interpolation/extrapolation • Evaluation across scales • Landscape level controls on biogeochem. • Understand cause of spatial variability • Emergent properties of landscapes

6. Why regional? Courtesy: Nic Saliendra

7. Why regional? • NEP (=-NEE) at 13 sites • Stand age matters • Ecosystem type matters • Is interannual variability coherent? • Are we sampling sufficient land cover types”?

8. Why data assimilation? • Meteorological, ecosystem, and parameter variability hard to observe/model • Data assimilation can help isolate model mechanisms responsible for spatial variability • Optimization across multiple types of data • Optimization across space

9. Why data assimilation? • Old way: • Make a model • Guess some parameters • Compare to data • Publish the best comparisons • Attribute discrepancies to error • Be happy

10. Why data assimilation? • New way: • Constrain model(s) with observations • Find where model or parameters cannot explain observations • Learn something about fundamental interactions • Publish the discrepancies and knowledge gained • Work harder, be slightly less happy, but generate more knowledge

12. Back to those stats… [A|B] = [AB] / [B] [P|D] = ( [D|P] [P] ) / [D] (parameters given data) = [ (data given parameters)× (parameters) ] / (data) Posterior = (Likelihood x Prior) / Normalizing Constraint

13. For the visually minded • D Nychka, NCAR

14. Some case studies • Prediction • Up and down scaling • Regional evaluation • Interannual variability • Forest disturbance and succession

15. Regional Prediction

16. Our tower is bigger…

17. Is there a prediction signal?

18. Sipnet • A “simplified” model of ecosystem carbon / water and land-atmosphere interaction • Minimal number of parameters • Driven by meteorological forcing • Still has >60 parameters • Braswell et al., 2005, GCB • Sacks et al., 2006, GCB added snow • Zobitz et al., 2008

19. Parameter estimation • MCMC is an optimizing method to minimize model-data mismatch • Quasi-random walk through parameter space (Metropolis) • Prior parameters distribution needed • Start at many random places (Chains) in prior parameter space • Move “downhill” to minima in model-data RMS by randomly changing a parameter from current value to a nearby value • Avoid local minima by occasionally performing “uphill” moves in proportion to maximum likelihood of accepted point • Use simulated annealing to tune parameter space exploration • Pick best chain and continue space exploration • Requires ~500,000 model iterations (chain exploration, spin-up, sampling) • End result – “best” parameter set and confidence intervals (from all the iterations) • NEE, Latent Heat Flux (LE), Sensible Heat Flux (H), soil moisture can all be used • Nighttime NEE good measure of respiration, maybe H? • Daytime NEE, LE good measures of photosynthesis • SipNET is fast (<10 ms year-1), so good for MCMC (4 hours for 7yr WLEF) • Based on PNET ecosystem model • Driven by climate, parameters and initial carbon pools • Trivially parallelizable (needs to be done, though)

20. Goldilocks effect…

21. 2 years = 7 years 1997 1998 1999 2000 2001 2002 2003 2004 2005

22. Regional futures

23. Regional futures

24. Upscaling and Downscaling

25. So many towers…

26. …so much variability

27. Simple comparisons… Desai et al, 2008, Ag For Met

28. …don’t work

29. We need to do better • Lots of flux towers (how many?) • Lots of cover types • A very simple model • Have to think about the tall tower flux, too • What does it sample?

30. Multi-tower synthesis aggregation with large number of towers (12) in same climate space • towers mapped to cover/age types • parameter optimization with minimal 2 equation model

32. Heterogeneous footprint

33. Tall tower downscaling • Wang et al., 2006

34. Scaling evaluation • Desai et al., 2008

35. Scaling sensitivity

36. Now we can wildly extrapolate • Take 17 towers • Fill the met data • Use a simple model to estimate parameters for each tower using MCMC • Apply parameters to other region meteorology data • Scale to region by cover/age class

37. Another simple(r) model • No carbon pools • GPP model driven by LAI, PAR, Air temp, VPD, Precip • LAI model driven by GDD (leaf on) and soil temp (leaf off) • 3 pool ER, driven by Soil temp and GPP • 19 parameters, fix 3

40. 20 yr regional NEE • Cover types + • Age structure + • Parameters • Forcing for a lake organic carbon input model

41. Regional scale evaluation

42. Top down and bottom up

43. IAV not modeled well

44. Region Interannual variability

45. Ricciuto et al.

46. Ricciuto et al.

48. IAV • Does growing season start explain IAV? • Can a very simple model be constructed to explain IAV? • Hypothesis: growing season length explains IAV • Can we make a cost function more attuned to IAV? • Hypothesis: MCMC overfits to hourly data

49. New cost function • Original log likelihood computes sum of squared difference at hourly • What if we also added monthly and annual squared differences to this likelihood? • Have to scale these less frequent values