16:00 online and NORCE 3rd floor conference room.
Talk by Matthias Morzfeld of Scripps (UCSD).
Abstract
Covariance localization is a key feature of ensemble Kalman filters (EnKF) in high-dimensional problems. Covariance localization typically relies on the assumption that correlations decay with distance. This assumption may no longer be valid when EnKFs are used in new application areas (e.g., Space Weather, geomagnetic data assimilation) or with more complex models that feature a more delicate correlation structure. We present a new method for covariance estimation which we call NICE – noise informed covariance estimation. NICE corrects ensemble estimates of correlations based on an estimated noise level within the correlations. NICE is adaptive (tuning-free), computationally easy to use and is robustly applicable in many test problems we tried. NICE also compares favorably to many other covariance corrections schemes taken from the statistical literature or the EnKF community.
