Plant trait meta-analysis and flux data assimilation constraints on parameterizations of ecosystem models

Background/Question/Methods

Probabilistic assessment and prediction of ecosystem functioning using mechanistic models is important for the advancement and communication of ecosystem science. We have developed the Predictive Ecosystem Carbon Analyzer (PECAn) to support the integration of data and models. PECAn is an open source workflow management tool that facilitates the integration of data into probabilistic forecasts.

Here, we describe the development of the PECAn data assimilation module and how it compliments the existing components of PECAn. The result is an integrated approach to constraining parameters that uses both trait data and observations of ecosystem level functions. PECAn quantitatively summarizes trait-level data using Bayesian meta-analysis to generate probability distributions that serve as the priors in the data assimilation step. This approach builds on existing data assimilation methods that begin with non-informative priors. In addition, the use of a model emulation approach instead of a MCMC approach to data assimilation reduces computational requirements of this step.

     These methods have been applied to a Switchgrass monoculture at the University of Illinois' Energy Farm. We present an example in which a diverse set of field observations, including remote sensing, soil respiration, and eddy-covariance measurements of carbon, moisture, and energy fluxes, are combined to constrain model parameters.

Results/Conclusions

The dual constraint on trait parameters by both meta-analysis and data assimilation reduced the uncertainty in parameter values more than using either method alone. Data assimilation alone resulted in greater parameter precision than meta-analysis alone but less accurate parameter estimation (i.e. less biologically realistic parameter estimates). Constraining parameters prior to data assimilation removed problems with equifinality and parameter identifiability that occured when non-informative priors are used. Estimating the likelihood using model emulation reduced the number of model runs required to estimate the parameter distributions from > 10⁶ often used with MCMC to < 1000. This exponential reduction in computational cost makes data assimilation with computationally intensive models feasible; the computational time required to use an MCMC approach with the ED2 model would be years, but PECAn completed the analysis in a few hours. Further work will include the implementation of a real-time data-assimilation and forecasting system.