Assimilation of Radar Data for Thunderstorm Prediction with Ensemble Kalman Filter: A Real Case Study

Recent OSSE studies have shown that the state of convective storms can be analyzed very accurately from simulated radar data with the ensemble Kalman filter (EnKF) method. However, the testing of the EnKF with real data, especially at the thunderstorm scale, remains very limited. For the assimilation of real data using the EnKF method, the presence of model error poses probably the biggest challenge. Without properly accounting for the model error, there tends to be a major discrepancy between the spread of the background ensemble forecast and the true background error. The background error covariance estimated from the ensemble can therefore be unrealizable.

We apply here the EnKF method to a tornadic thunderstorm case that occurred on May 29, 2004 near the Oklahoma City. A long lasting supercell produced sixteen tornadoes over a period of several hours. Our recent results assimilating observed radar data for this case, in which the storm environment is defined by single sounding and the ARPS model is used in a cloud model mode, is encouraging. Radial velocity and reflectivity fields at the end of 1 h assimilation window match observations well. The flow fields show dynamically consistent patterns typical of supercell storms, including strong mid-level rotation and hock echo. Predicted storm maintained supercell characteristics for more than 1 hour, but was generally weaker and propagated too fast. Errors in the forecast model and the storm environment defined by the single sounding are believed to the main culprit.

In this study, the initial storm environment will be inhomogeneous and provided by a 3DVAR scheme that includes all types of observations. A complex cloud analysis procedure is used to initialize the initial thunderstorms. The prediction model will include full physics. The full physics package and properly analyzed storm environment are expected to improve the EnKF analysis of the thunderstorms as well as the subsequent forecast. The issues of model error will be also addressed in this study.