P1.44 The Issue of Data Density and Frequency with EnKF Radar Data Assimilation in a Compressible Nonhydrostatic NWP Model

1. Introduction Since its first introduction by Evensen (1994), the ensemble Kalman filter (EnKF) technique for data assimilation has received much attention. Rather than solving the equation for the time evolution of the probability density function of model state, the EnKF methods apply the Monte Carlo method to estimate the forecast error statistics. A large ensemble of model states are integrated forward in time using the dynamic equations , the moments of the probability density function are then calculated from this ensemble for different times (Evensen 2003). Recently, EnKF was applied to the assimilation of simulated Doppler radar data for modeled convective 2007) with great successes. But the application to real radar data was not very elegant (Dowell et al. 2004; Tong and Xue 2007). One of the advantages of EnKF method over variational method is that it can dynamically evolve the background error covariances throughout the assimilation cycles, thereby providing valuable uncertainty information on both analysis and forecast. Recently, Caya et al. (2005) showed that with simulated radar data, the EnKF method can outperform a similarly configured 4DVAR scheme after the first few assimilation cycles. When combined with an existing ensemble forecast system (operational ensemble forecasting system is usually run at a lower resolution compared to the operational deterministic forecast), the EnKF method can provide quality analyses with a relatively small incremental cost compared to a 4DVAR system that requires repeated integrations of the forward prediction model and its adjoint. Same as 4DVAR, the overall computational cost of ensemble-based assimilation methods is significant because of the need for running an ensemble of forecast and analysis of nontrivial sizes (usually a few tens to a few hundreds), especially when high-density data are involved and when the ensemble of all forecasts is run at high resolutions. One of the major sources of errors with the EnKF is the sampling error associated with the limited ensemble size. A larger ensemble helps improve the background error covariance estimation, but incurs