Extended Abstract (68K)6.2
An Ensemble Kalman Smoother for Reanalysis
Jeffrey S. Whitaker, NOAA/ERL/CDC, Boulder, CO; and G. P. Compo
Data assimilation is typically used to generate initial conditions for numerical weather forecasts. Therefore, each analysis is based upon only current and past observations. However, when producing a retrospective 'reanalysis', one is free to use all available observations, including those data collected after the analysis time.
A Kalman smoother is a direct generalization of the Kalman filter which incorporates observations both before and after the analysis time. Here we introduce the ensemble Kalman smoother (EnKS), which applies recent advances in the field of ensemble filtering to the fixed-lag Kalman smoother proposed by Cohn and collaborators. The EnKS uses Monte-Carlo estimates of forecast-analysis error covariances needed to compute the Kalman smoother gain matrix. It is applied iteratively to a time series of observations, the first iteration is equivalent to an ensemble Kalman filter analysis which only utilizes observations taken up to and including the analysis time. The n-th iteration utilizes observations taken n observing times past the analysis time. Only the first iteration requires the integration of a forecast model.
We present results with the EnKS in a hierarchy of idealized models, and examine the impact of observations taken after the analysis time as a function of 1) the observing network, 2) the ensemble size, and 3) the amplitude of the model error. A three-dimensional variational assimilation (3DVar) using only observations taken prior to the analysis time is used as a baseline for comparison. The EnKS is found to improve upon the baseline most when the ensemble is large, the model is accurate and the observations are sparse. In particular, when only surface observations are available, the EnKS is able to produce a fairly accurate upper-tropospheric analysis, a situation in which 3DVar fails miserably. The potential for using the EnKS for historical reanalysis is discussed, with particular emphasis on reanalysis before the radiosonde era (pre-1948).
Session 6, Ensembles and data assimilation
Thursday, 17 January 2002, 8:45 AM-1:30 PM
Previous paper Next paper