Two convectionpermitting numerical experiments were conducted with the WRF model to examine the impact of assimilating sounding data by using an EnKF method for deterministic prediction of a torrential rainfall event over southern China during 15-16 May 2013. The initial and lateral boundary conditions were based on the NCEP GFS 1°×1° 60 h gridded forecast data which were available every 6 h from 08:00 BT 14 May to 20:00 BT 16 May 2013. The two experiments, NODA and DA, differed only in the initial conditions: while NODA was initialized from the NCEP GFS data at 08:00 BT 14 May, DA was from an ensemble mean of 30 analysismembers at 08:00 BT 15 May, which was generated using the WRFEnKF system with conventional sounding data at 20:00 BT 14 May, 02:00 BT and 08:00 BT 15 May assimilated. The results show that, compared to NODA, not only the initial conditions of DA are much closer to the observed fields, but also the DA predicted physical parameters are improved in terms of both biases and rootmeansquare errors, leading to a more accurate prediction of the location and magnitude of precipitation from DA.