State, Parameter and Noise Estimation for a Coupled Ocean-Atmosphere Model

T. M. Chin, J. B. Jewell, & M. J. Turmon

Jet Propulsion Laboratory and California Institute of Technology

Data assimilation practiced today is mostly based on linear estimation theory. In particular, a Gaussian distribution of estimation errors and a quasi-linear treatment of covariance dynamics have been assumed and practiced. This does not take account of the weather and climatic "regimes" (favored state values) that have been recognized both theoretically (e.g., Lorenz 1963) and empirically. A fundamental problem is that the Gaussian distribution is unimodal, while weather and climate often display multi-modal characteristics known also as dynamic regimes. We have applied Monte Carlo numerical techniques to some low-order problems without assuming a Gaussian distribution and quasi-linear dynamics. First, we have extended the well-known assimilation technique of "ensemble Kalman filter" to a smoothing algorithm, referred to as the "resampled particle smoother" (RPS), based on theoretical developments by Kitagawa (1996) and others. The particle smoother is observed to yield more accurate state estimates than the corresponding filtering algorithm.

Last Update: March 25, 2005