Premium
Ensemble Kalman filtering
Author(s) -
Houtekamer P. L.,
Mitchell Herschel L.
Publication year - 2005
Publication title -
quarterly journal of the royal meteorological society
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.744
H-Index - 143
eISSN - 1477-870X
pISSN - 0035-9009
DOI - 10.1256/qj.05.135
Subject(s) - ensemble kalman filter , data assimilation , kalman filter , ensemble forecasting , ensemble learning , dimension (graph theory) , computer science , interpolation (computer graphics) , range (aeronautics) , ensemble average , filter (signal processing) , mathematics , meteorology , algorithm , extended kalman filter , artificial intelligence , physics , climatology , geology , aerospace engineering , pure mathematics , engineering , computer vision , motion (physics)
An ensemble Kalman filter (EnKF) has been implemented at the Canadian Meteorological Centre to provide an ensemble of initial conditions for the medium‐range ensemble prediction system. This demonstrates that the EnKF can be used for operational atmospheric data assimilation. We show how the EnKF relates to the Kalman filter. In particular, to make the ensemble approximation feasible, we have to use a fairly small ensemble with many less members than either the number of model coordinates, or the number of independent observations, or the (unknown) dimension of the dynamical system. To nevertheless obtain good results, we must (i) counter the tendency of the ensemble spread to underestimate the true error, and (ii) localize the ensemble covariances. The localization is severe and leads to imbalance in the initial conditions. The operational EnKF is used to investigate to what extent our system respects the underlying hypotheses of both the Kalman filter and its ensemble approximation. In particular, we quantify the imbalance in the initial conditions and the magnitude of the model‐error component. The occurrence of imbalance constrains the ways in which time interpolation can be performed and in which parametrized model error can be added. With this study we hope to obtain and provide guidance for further improvements to the EnKF. Copyright © 2005 Royal Meteorological Society