A data-driven method for improving the correlation estimation in serial Ensemble Kalman filter

Michele De La Chevrotiere* and John Harlim
Penn State University, Mathematics Department

Ensemble Kalman filters (EnKF) with small ensemble size tend to induce spurious long-range correlations in the ensemble approximation of the model’s covariance. The typical approach to this long standing issue is by using space localization techniques that effectively reduces the spurious correlations. Many such techniques have been proposed, for instance with the tapering functions of Furrer and Bengtsson (2007), the Gaspari and Cohn localization functions (1999), and other distance-based localization functions for multiphase flow (see, for instance, Chen and Oliver, 2010). While these techniques have been very useful, they require exhaustive tuning and present challenges when applied to nonlinear observations. Recently, Anderson and Lei (2013) have introduced an approach based on empirical localization functions (ELF) that requires almost no tuning. However, ELF are constructed in stages and have limitations when applied to large atmospheric models. Motivated by this approach, we present a data-driven method for improving the sample correlation estimation in the EnKF when small ensemble size is used. In particular, we find a linear map that takes the poorly estimated sample correlation function in each EnKF cycle and transforms it into a sample of improved correlation matrix. This talk will present an overview of the method and some preliminary results.

*email: mdelachev@psu.edu
*Preference: Oral