Uncertainty management in coupled physical-biological lower trophic level ocean ecosystem models

Abstract

Lower trophic level (LTL) ocean ecosystem models are important tools for understanding ocean biogeochemical variability and its role in Earth's climate system. These models are often replete with parameters that cannot be well constrained by the sparse observational data available. LTL ocean ecosystem model parameter estimation is examined from a probabilistic perspective, using a Bayesian hierarchical model (BHM), in the coastal Gulf of Alaska (CGOA) domain that benefits from ocean station observations obtained in repeated US GLOBEC cruises. Data entering the BHM include daily average SeaWiFS satellite estimates of surface chlorophyll and GLOBEC observations of nutrient and phytoplankton profiles at inner and outer shelf stations on the Seward Line. The final form of the BHM process model component is comprised of a discrete version of the Nutrient-Phytoplankton-Zooplankton-Detritus LTL ecosystem model equations augmented to address iron limitation in the CGOA (i.e., NPZDFe), and including a vertical diffusion term to constrain the timing of the phytoplankton bloom in spring. Even in the relatively data-rich GLOBEC context, parameter estimation in the BHM requires guidance from a suite of calculations in a coupled physical-biological deterministic model-the Regional Ocean Model System coupled to an NPZDFe component (ROMS-NPZDFe). ROMS-NPZDFe simulations are used to: (1) validate the BHM formulation, (2) separate BHM limitations due to sampling from those due to LTL model approximations, and (3) obtain output distributions for zooplankton grazing rate and phytoplankton nutrient uptake rate using GLOBEC and SeaWiFS data for 2001. Uncertainty is evident from the spreads in output distributions for model parameters in the BHM. Experiments driven by simulated data from ROMS-NPZDFe helped to optimize the utility of GLOBEC observations for LTL ocean ecosystem model parameter estimation, given ever-present uncertainty issues. The ROMS-NPZDFe simulations are also used to build Bayesian statistical models as surrogates for the deterministic model. Two applications are briefly described. One estimates output distributions for selected ocean ecosystem parameters while accounting for spatial variability across the GLOBEC stations in the CGOA. A second application assimilates SeaWiFS data and simulated data from a ROMS-NPZDFe control run for 2002 to estimate complete fields of surface phytoplankton concentration, with associated spatial and temporal uncertainties. © 2013 by The Oceanography Society. All rights reserved.