Two promising techniques for assimilating radar data into numerical models are 4D Variational data assimilation (4DVar) and the Ensemble Kalman Filter (EnKF). Both of these techniques use a time series of data and a model to find an analysis which is the best fit to the data and a background field (given error covariances for the data and background fields). In the limit of a perfect model and linear dynamics, 4DVar and the Kalman filter give similar results at the end of the assimilation window. However, in a real situation, with an imperfect model, nonlinear dynamics and where it is impossible to determine the full background error covariance matrix, the two techniques are expected to give different results. In this paper, we will examine these differences by applying the two techniques to radar observations of a convective storm.

The case we have chosen for this study is the 17 May 1981 Arcadia, OK, supercell which was observed by two Doppler radars spaced approximately 40 km apart. We assimilate data from one radar and verify the retrieved wind field with observations from the second radar. We also use dual Doppler analyses to verify both the vertical velocity and crossbeam velocity retrieved using single Doppler observations. A number of assimilation experiments have been conducted using a time series of single Doppler observations. Short term forecasts have also been conducted using the retrieved fields as initial conditions.

This work is ongoing and results obtained so far will be presented at the Conference. In general terms, both techniques are able to retrieve the three-dimensional flow field with reasonable accuracy. The r.m.s. differences between the retrieved and dual-Doppler-calculated crossbeam velocity fields are of the order of 5 m/s while the maximum vertical velocity retrieved is within 20-30% of its dual-Doppler counterpart. There is a tendency for the 4DVar technique to perform better in the early part of assimilation time window, while EnKF does better latter in the window due its improved calculation of the background error covariances.