Impact of Removing Covariance Localization in an Ensemble Kalman Filter: Experiments with 10 240 Members Using an Intermediate AGCM

Abstract

Abstract The ensemble Kalman filter (EnKF) with high-dimensional geophysical systems usually employs up to 100 ensemble members and requires covariance localization to reduce the sampling error in the forecast error covariance between distant locations. The authors’ previous work pioneered implementation of an EnKF with a large ensemble of up to 10 240 members, but this method required application of a relatively broad covariance localization to avoid memory overflow. This study modified the EnKF code to save memory and enabled for the first time the removal of completely covariance localization with an intermediate AGCM. Using the large sample size, this study aims to investigate the analysis and forecast accuracy, as well as the impact of covariance localization when the sampling error is small. A series of 60-day data assimilation cycle experiments with different localization scales are performed under the perfect model scenario to investigate the pure impact of covariance localization. The results show that the analysis and 7-day forecasts are much improved by removing covariance localization and that the long-range covariance between distant locations plays a key role. The eigenvectors of the background error covariance matrix based on the 10 240 ensemble members are explicitly computed and reveal long-range structures.