Jan Frydendall, DTU

Data Assimilation in Marine Models

In this presentation the topic will be Data Assimilation in Marine
Models.  The Ensemble Kalman Filter (EnKF) is the de-facto standard of
modern Data Assimilation in marine models.  The EnKF is, however, a
B.L.U.E. estimator and is often an unfit filter for the deterministic
state spaces models. These state space models have often a non-linear
dynamics that drives them. These non-linear effects most often only
have an effect on smaller parts of the state space and can therefore
be neglected for the most parts. However, some times it could be of
interest to information about some for these non-linear effects,
especially if these small parts of the model shall be used as input to
another model, for examples Eco-system models. For example the
temperature profile around a pycnocline (layer of separation) in the
Øresundstrait. Or if one would like make assessments on the effects
of bridge pillars on the surroundings before a bridge is
constructed. Nonetheless the EnKF is s till being used since is a very
robust filter.

Particle filter s has been suggested as a solution to
such non-linear cases. However, the importance sampling filters have
serious problem with the high dimensions of the state space
models. Since on average a marine models has a state space dimension
of 10,000~100,000. This is in most cases infeasible for any
particle filter. Localizations of the state space is a possibility and
has been suggested to reduce the dimensions of the state space in the
localized regions.

In the talk a will discuss the definition of data assimilation in
marine models, and outline some of the main problems in data
assimilation. There after I will talk about the Ensemble Kalman Filter
and show some results that can be obtained with the EnKF in marine
models. I will also address the aspect of data assimilation with
particle filters and show some preliminary results obtained with the
bootstrap filter and some results of localization.