Extended ensemble Kalman filters for data assimilation in hierarchical state-space models

Matthias Katzfuss* (1) and Jonathan R. Stroud (2), Christopher K. Wikle (3)
(1) Department of Statistics, Texas A&M University
(2) McDonough School of Business, Georgetown University
(3) Department of Statistics, University of Missouri

The ensemble Kalman filter (EnKF) is a computational technique for approximate inference on the state vector in state-space models. It has been successfully used in many real-world data-assimilation problems with very high dimensions. However, the EnKF assumes the state-space model to be fully known and to be at least approximately linear and Gaussian. Here, we consider a broader class of hierarchical state-space models, which include two additional “layers”: The parameter layer allows handling of unknown parameters that cannot be easily included in the state vector, while the observation layer can be used to model non-Gaussian observations. We propose a general class of extended ensemble Kalman filters and smoothers for (approximate) Bayesian inference in our hierarchical state-space models. We highlight several interesting examples, such as a robust EnKF for heavy-tailed observations, and assimilation of hourly rainfall amounts.

*email: katzfuss@gmail.com
*Preference: Poster