Previously we used the Predictive Ecosystem Analyzer (PEcAn) to assess the contribution of different model parameters to the uncertainty of the Ecosystem Demography model v2 (ED), a demographically-enabled terrestrial biosphere model, across a selection of North American biomes (Dietze et al., JGR-G, 2014). Through this analysis we identified key research priorities for model improvement, but it was unclear if these results were model and/or biome specific. In this work we expand on that previous study by focusing on the Barro Colorado Island (BCI) site using updated PEcAn tools and a range of models that vary in spatial resolution and complexity. Local driver and parameter data from the DOE-funded NGEE Tropics project are used in addition to values from the BETYdb and TRY trait databases. Parameter uncertainties are assessed using an updated multivariate version of PEcAn’s Hierarchical Bayesian meta-analysis that is used to provide model parameter posterior distributions. In addition, a sensitivity analyses across the model suite is performed to assess sensitivity of model outputs to parameters. Finally, parameter uncertainties and model sensitivities are combined to evaluate the fractional contribution of each parameter to the predictive uncertainty for a specific variable.
Results/Conclusions
The parameters to which NPP was most sensitive varied considerably among models. For example, in the SIPNET model NPP was most sensitive to the optimal photosynthetic temperature, while in the DALEC model it was most sensitive to specific leaf area (SLA). Taking a general look at which processes these parameters belong to shows that overall the models are most sensitive to canopy level processes. We used parameter prior distributions from BETYdb that were not well constrained but by using PEcAn’s automated tools we can automatically perform our analysis again using parameters from the NGEE tropics team and TRY database, that are BCI specific, allowing us to use more constrained parameter estimates. Integrating this data with models will allow us to make conclusions about the relative parameter uncertainty associated with different process representations amongst models. Ultimately, our multi-model analysis can be used to direct future experimental designs and field data collection activities, by illuminating parameters comprising model PFTs that need further data constraint, and highlight which process representations must be better understood.