A sensitivity study of radar location in data assimilation using the ensemble Kalman filter

The sensitivity of radar location in the ensemble Kalman filter (EnKF) data assimilation is studied. A suite of experiments has been conducted to simulate an idealized supercell storm case with assimilation of observations from radars located at various locations. There are eight radars used in these experiments. These radars are homogeneously placed around the simulated domain so that radar observations at all quadrants can be taken into account. Assimilations of observations from single radar and from two radars are investigated respectively. Results of the single radar data assimilation show that the radar location has obvious impacts on the first several cycles of EnKF data assimilation and these impacts become weak after more than ten cycles. Radars that can cause large analysis errors are found to be located at the north and south of the model domain. The lines connecting the center of the model domain and the locations of the two radars are orientated perpendicular to the moving direction of the storm (which is also the environmental flow direction). Results of the experiments with assimilation of observations from two radars show that the EnKF performs well when both radars are located at the diagonal of the simulated domain (the angle between the environmental flow and the lines connecting the radar locations and the storm center is 45°) and when the line between the domain center and one radar is orientated perpendicular to that between the domain center and the other radar. However, when both radars are located at the north, south, east or west of the simulated domain, the performance of EnKF with observations from two radars becomes worse in the early cycling stage. Results of the deterministic forecast after a short-term data assimilation show that the large analysis errors caused by radar location can further deteriorate the performance of subsequent forecasts. Those large analysis errors are caused by the poor error covariance driven by the inaccurate initial condition. With the improvement of background after more cycles, the covariance becomes better and the difference between experiments caused by using observations from different radars becomes smaller. Based on the above result, the iterative EnSRF (iEnSRF) is introduced. Results show that the performance of data assimilation in early cycles and the subsequent forecast are significantly improved when iEnSRF is employed.