Calibration of the NEMURO nutrient–phytoplankton–zooplankton food web model to a coastal ecosystem: Evaluation of an automated calibration approach

A one spatial-box version of the NEMURO oceanic lower trophic food web model was applied to a coastal upwelling environment typified by West Coast Vancouver Island. We used both ad hoc calibration and the automatic calibration program PEST. NEMURO was first calibrated to 1 year of monthly field data using the usual ad hoc approach of trial and error changes to 18 candidate parameters. Four PEST calibrations were then performed. The first three PEST calibrations used model predictions in year 10 from the ad hoc calibration as data in a twin experiment design; the fourth PEST calibration repeated the ad hoc procedure by having PEST calibrate NEMURO to the field data. When provided with ad hoc calibration model predictions as data, PEST accurately recovered the known 18 parameter values, even when small and large phytoplankton were lumped into total phytoplankton. When 57 parameters were allowed to vary PEST-estimated reasonable values for all 57 parameters, but they differed from the ad hoc calibrated values. However, when applied to the field data, PEST-estimated parameter values that differed greatly from the ad hoc values. The PEST calibration fitted some of the field data better than the ad hoc calibration but at the cost of unequal small and large phytoplankton concentrations. Thus, with proper and careful implementation, PEST offers a viable approach for objective calibration of NEMURO to site-specific monitoring data. We recommend that automatic calibration methods, such as PEST, be used for application of the NEMURO model to new locations. When the field data allow for specification of time series for each phytoplankton and zooplankton state variable, PEST will provide an objective, defensible, and repeatable way to calibrate the many parameters of the NEMURO model. If the available data are insufficient for specification of each state variable, then ad hoc calibration will likely be needed to allow for inclusion of qualitative decisions about model fit. Use of PEST in this situation will provide better understanding of the data-model mis-matches and will provide an alternative calibration to the necessary, but subjective, ad hoc calibration. Comparison of the ad hoc and PEST calibrations (even if unsuccessful) will help in the interpretation of the ad hoc calibration. Robust parameter estimation by any method depends on the quality and consistency of the calibration dataset.