Ensemble Filters for Atmosphere and Ocean Data Assimilation

Modern data assimilation for the atmosphere and ocean combines information from observations and a prediction model to produce an estimate of the state of the physical system. Data assimilation is used not only to generate 'analyses' of the system state, but also to produce initial conditions for prediction models. It can also be used to improve both the observing system and the prediction models.

Ensemble filtering algorithms for data assimilation have recently begun to mount a challenge to variational methods as candidates for next generation prediction systems. A simple derivation of several varieties of ensemble filters as Monte Carlo approximations to the solutions of a non-linear filtering problem is presented. Most of the ensemble methods in use can be derived as a one-dimensional filter followed by a sequence of linear regressions. Selected results from applications in both low-order models and large operational prediction models demonstrate the power of these methods.