A Bayesian parameter estimation method applied to a marine ecosystem model for the coastal Gulf of Alaska

• Published in: Ecological modelling Vol. 258; pp. 122 - 133
• Main Authors: Fiechter, J., Herbei, R., Leeds, W., Brown, J., Milliff, R., Wikle, C., Moore, A., Powell, T.
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier B.V 01.06.2013; Elsevier
• Subjects: Adaptive Metropolis–Hasting; algorithms; Animal and plant ecology; Animal, plant and microbial ecology; Bayesian sampling; Biological and medical sciences; chlorophyll; ecosystems; Fundamental and applied biological sciences. Psychology; General aspects; General aspects. Techniques; grazing; Gulf of Alaska; Marine ecosystem model; Methods and techniques (sampling, tagging, trapping, modelling...); nitrates; Parameter estimation; phytoplankton; Posterior distributions; remote sensing; Sea water ecosystems; Synecology; trophic relationships; zooplankton; Gulf of Alaska; Pacific Ocean; North Pacific; Marine ecosystem model; Parameter estimation; Bayesian sampling; Adaptive Metropolis–Hasting; Gulf of Alaska; Posterior distributions; Marine environment; Ecosystem; Models; Sampling; Adaptive Metropolis―Hasting; Adaptation
• ISSN: 0304-3800; 1872-7026
• DOI: 10.1016/j.ecolmodel.2013.03.003

•Implementation of state-of-the-art adaptive algorithm to facilitate Bayesian sampling.•Bayesian estimates of posterior distributions for marine ecosystem model parameters.•Impact of observation sampling and availability on parameter posterior distributions.•Assessment of Bayesian method to estimate of multiple correlated model parameters. The present study describes a state-of-the-art methodology based on an adaptive Metropolis–Hastings algorithm to facilitate efficient Bayesian sampling for realistic lower trophic level (LTL) marine ecosystem models. The main objective is to explore the ability to differentiate between biological parameters that can learn from observations and those that cannot. The Bayesian approach is applied to the northwestern coastal Gulf of Alaska region and uses both synthetic and actual (in situ and remotely sensed) observations. LTL ecosystem dynamics in the Bayesian framework are described by a process model consisting of a 1-dimensional Nutrient–Phytoplankton–Zooplankton–Detritus formulation with iron limitation (NPZDFe) and vertical mixing. The results illustrate the ability to determine parameter posterior distributions for fundamental biological rates, such as maximum phytoplankton growth or zooplankton grazing. By using various observational platforms as data stage inputs, the results also demonstrate the impact of spatial and temporal sampling on parameter posterior distributions, as well as the benefits of having concurrent measurements for two or more state variables of the process model (e.g., chlorophyll and nitrate concentrations). Extending the method to multiple parameters is non-trivial, as posterior distributions become impacted by correlated and/or disproportionate contributions for certain model parameters. Controlled experiments with “near perfect data” were useful to characterize parameter identifiability based on information content in the BHM data stage inputs, as well as to separate uncertainties due to sampling issues vs. uncertain ecosystem process interpretation.