| Summary: | Better methods are required to assess the skill or uncertainty of plankton model predictions. A method is presented which combines cross-validation with simulated repeat samplings of the data (Monte Carlo simulation), in order to robustly estimate uncertainty in predictions beyond the calibration data ('extra-sample'). The method is applied to compare two bulk models of chlorophyll on Georges Bank using the GLOBEC data set, accounting for data and forcing errors as well as prior uncertainty in all model parameters and initial conditions. The first model is a simple interpolation of chlorophyll data ('inductive' model), and serves as a baseline of predictive skill. The second is a simple process model forced by interannually-variable nutrient and mesozooplankton mean fields. Uncertainty in the process model forcings severely increases the extra-sample prediction variance (over repeat experiments). Although the process model can reproduce some of the interannual chlorophyll variability via top-down control by mesozooplankton, other predictions are strongly biased, possibly due to neglected boundary fluxes of chlorophyll. As a result, the new skill metrics generally favour the inductive model. By contrast, a standard skill metric based on calibration data misfit incorrectly favours the process model, mainly due to the neglect of extra-sample prediction variance.
|
|---|
