Ensembles and Particle Filters for Ocean Data Assimilation
Abstract
It is our long-range goal to develop efficient methods for construction of error estimates based on the probability density functions of stochastic forecast models, and to apply those methods to the construction of practical data assimilation systems. We wish to understand in detail the mechanisms for propagation of uncertainty, and to investigate and devise new methods for quantifying the information content of forecasts and analyses. Practical models of the ocean and atmosphere have typical state dimensions of O(10(5) - 10(7) ), so direct calculation of the probability density function (pdf) is not practical. We must therefore apply Monte-Carlo methods, in which we draw ensembles, i.e., collections of samples drawn from the pdfs in question. In this context, the data assimilation problem becomes that of choosing an ensemble of points in state space drawn from the background, i.e., the "prior" pdf, and deriving a corresponding ensemble drawn from the "posterior" pdf, i.e., the pdf informed by observations. Data assimilation methods that use ensembles in the estimates of the prior and posterior pdfs are widely known as "particle methods," since each ensemble member can be considered as a particle in state space. Since our long-range objective is the evaluation of the evolution of the pdfs of model state spaces, informed by observation, we will necessarily be concerned with particle methods.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 30, 2010
- Accession Number
- ADA542455
Entities
People
- Robert N. Miller
Organizations
- Oregon State University