Work

Adaptive localisation of the ETKF by domain clustering of the forecast ensemble

Year: 2026
Type: preprint
Abstract: Ensemble data assimilation methods, such as the Ensemble Kalman Filter (EnKF), are inherently sensitive to sampling errors caused by the small size of the ensemble relative to the high dimension of geophysical models. To prevent spurious long-range correlations, localization techniques are applied to restrict the spatial influence of observations. However, traditional domain localization methods rely on fixed, predefined localization radii. This rigid partitioning does not account for the anisotropic error structures and the actual physical dynamics of the system.This paper introduces an adaptive domain localization strategy that incorporates the physical geometry of the ensemble at the time of analysis. The approach uses a K-Means clustering algorithm directly in the state space to group grid points that display similarities across the ensemble at the time of the analysis. To ensure spatial consistency and maintain a scale appropriate for the ensemble size, the resulting clusters are geometrically segmented to separate disconnected regions and resized if they exceed a maximum span threshold. Observations are then assigned to these data-driven clusters based on a minimum distance criterion to the cluster boundaries.Numerical experiments conducted on a quasi-geostrophic model simulating large-scale ocean circulation show that the clustering successfully tracks the formation and displacement of dynamic structures, such as vortices. The proposed framework effectively reduces the root mean square error (RMSE) while preserving the parallel computation advantages of the Local Ensemble Transform Kalman Filter (LETKF).
Source: HAL
Language: English
Open Access status: green
OpenAlex ID: W7150389982
Cites: 26
Cited by:
Related to: 10