Coherence of parameters governing NEE variability in eastern U. S Coherence of parameters governing NEE variability in eastern U.S. forests: A multisite data assimilation of eddy covariance data using TRIFFID. Daniel Ricciuto June 5th, 2006 ChEAS meeting IX
Motivation Terrestrial models are used to predict future fluxes. Large uncertainty, grows with time Generally not constrained by observations How can flux towers help? Friedlingstein et al. (in press)
Data assimilation and flux towers Previous work Braswell (2005), Sacks (2006): Parameter optimization using SipNET model Improves model estimates of NEE Captures seasonal cycle Interannual variability poorly modeled
Data assimilation and flux towers Ricciuto et al. (in press) Simple model: based on gap-filling routines, includes SWC dependence reproduces daily sums of NEE reasonably well (example month: Sep 1997) Also reproduces seasonal cycle of monthly NEE sums reasonably well
Data assimilation and flux towers Interannual variability in NEE sums poorly modeled, but does capture 1997-2001 difference. Nighttime NEE variability modeled reasonably well but model is biased low (because respiration parameters include daytime data) Daytime NEE modeled reasonably well but model biased high (because respiration parameters include nighttime data) 1997 2001
Multiple tower assimilation No published studies yet using data assimilation with multiple flux towers (others in press) Key questions: Can a single set of optimized model parameters: reproduce observed interannual variability? reproduce observed intersite variability? Are parameters coherent across space? Time? Model: Top-town representation of interactive foliage and flora including dynamics (TRIFFID)
The TRIFFID world: 5 PFTs Source: Hadley Centre (http://www.metoffice.com/research/hadleycentre/models/carbon_cycle/models_terrest.html)
Modified TRIFFID carbon cycle Climate input: Precip, PAR, Tair, RH Ra1 GPP1 GPP2 Ra2 Broadleaf PFT (H1, FRAC1, LAI1) Needleleaf PFT (H2, LAI2, FRAC2) wood wood RH Soil layer (Tsoil, CS, SWC) root root
Key model parameters Photosynthesis: Autotrophic respiration: Vmax, a, Tupp, Tlow, Q10VM, SWCopt, SWCdep Autotrophic respiration: Rdc, Rgc, Q10RD Heterotrophic respiration: Csoil Q10, SWCoptR, SWCdepR Phenology Toff, SWCoff, Lburst
Nonconvex problem Nonlinear systems often are nonconvex (multiple maxima in parameter space) Gradient-based optimization (e.g. Levenberg-Marquardt) misconverges Need a method that can find global (best) solution
Assimilation technique Stochastic Evolutionary Ranking Strategy (SRES) Global optimization method Stochastic – difficult to guarantee convergence Relatively fast method to find very good solution No parametric uncertainty (unlike MCMC) Method: Start with initial population (parameter sets) Evaluate goodness of fit (objective function) Select best-fitting members Mutate these members, repeat until convergence apparent Run twice for each tower, compare solutions
Method 5 eastern U.S. flux sites with long records (>= 5 years) Optimize each site individually with hourly data (23 params) Optimize all 5 sites jointly (single set of parameters) Soil carbon treated as separate fit parameter for each site
Results: seasonal cycle
Interannual variability fff
Results: Joint optimization
Conclusions Optimized TRIFFID model reproduces seasonal cycle at each eddy covariance site well Interannual variability poorly modeled. Why? Poor representation of hydrology (fit improved at WLEF when SWC data used… but don’t have everywhere) Do we need better pools / time lag effects? Intersite variability modeled somewhat well Soil carbon as only intersite variable: oversimplification
Future work Better representation of hydrology More data sources as constraints Hydrology where available Inventory data Manipulation experiments Run model over longer timescales Optimize parameters related to longer-term effects such as competition, CO2 fertilization Future predictions Couple to cheap GCMs, run ensembles of predictions to gauge uncertainty