Bayesian data assmilation experiments for numerical ocean models

Michael Dowd
Dalhousie University

Abstract
Data assimilation involves the blending of numerical models for nonlinear differential equation (DE) based systems with available non-Gaussian observations. The goal is produce optimal estimates of the system state and its parameters. A Bayesian approach to this time dependent estimation problem can be formulated based on the state space model. Sequential Monte Carlo methods provide for solutions in terms of samples which characterize the target distributions of interest. These rely on resampling (bootstrapping) approaches, or on MCMC; hybrid and approximate approaches are also available. The design of sequential data assimilation experiments involves considerations of which of the candidate methodologies to use, as well as the size of the samples required to appropriately characterize the target distributions of interest.

Towards this end, I consider some computer experiments for the identification of effective sequential data assimilation schemes. These are focused on both the computational requirements and convergence of the candidate methods. Assessment of the dependence structure amongst the prognostic variables is also undertaken. The specific system under consideration is governed by nonlinear Des describing ocean biogeochemistry. Observations are taken from a recently established coastal ocean observatory in Lunenburg, NS, Canada. The motivations for these experiments are to design a biogeochemical forecasting system for use in an operational context as part of a comprehensive regional ocean-atmosphere prediction system currently under development.

Back to agenda