Ecosystem model complexity versus physical forcing: Quantification of their relative impact with assimilated Arabian Sea data

In the wake of recent large-scale observational programs such as Joint Global Ocean Flux Study (JGOFS) and Global Ocean Ecosystems Dynamics (GLOBEC), a number of models have been developed to simulate biogeochemical cycling in various oceanographic regions; however, few quantitative comparisons of these models have been made. In order to assess critically which ecosystem structures and model formulations are best able to simulate observed biogeochemical cycling, three fundamentally different ecosystem models of varying complexity are applied within a consistent one-dimensional (1-D) framework at 15.5°N, 61.5°E in the Arabian Sea. Each model is forced by three different sets of physical forcing fields: two of these are derived from the solutions of different 3-D physical models, while the third is derived from moored observations. In situ concentrations of plankton and nitrogenous nutrients, and rates of production and export measured during the US JGOFS Arabian Sea Expedition are used for assimilation and evaluation. After objectively optimizing each model, their performance is quantitatively compared to assess which model structure best represents the fundamental underlying biogeochemical processes. Results are highly sensitive to the number of parameters optimized for each ecosystem model. A set of cross-validation experiments designed to assess predictive capability demonstrates that when too many parameters are allowed to vary, the more complex models are unable to reproduce any unassimilated data, suggesting that under these conditions the models have little predictive skill. Optimizing only an objectively and systematically selected subset of uncorrelated model parameters minimizes this problem. The method developed to accomplish this parameter selection is presented. After this optimization method is applied, all three models behave similarly, implying that the additional complexity of a multiple size-class model may not be advantageous. Furthermore, a change in physical model (mixed-layer depth and vertical velocity fields) typically produces a far greater change in biogeochemical model response than does a change in ecosystem model complexity, highlighting the fact that biogeochemical variability is largely determined by the physical environment both in situ and in the model domain.