In this study, the impact of rainfall assimilation on the forecasts of convective rainfall over the mountainous areas in the southwestern United States is investigated. The rainfall is derived from the U.S. Weather Surveillance Radar--1988 Doppler (WSR-88D) radar network, and the fifth-generation Mesoscale Model (MM5) Four-Dimensional Variational (4DVAR) system is employed in the study. We evaluate the rainfall assimilation skill through two rainstorm events (5--6 August and 11--12 September 2002) that occurred over the southwestern United States in 2002. A series of experiments for the two cases is conducted. The results show that the minimization process in the 4DVAR is sensitive to the length of assimilation window and error variance in the observation data. Assimilation of rainfall can produce a better short-range precipitation forecast. However, the time range of improved forecasts is limited to about 15 hours with the model resolution of 20 km. It is indicated that rainfall assimilation produces more realistic moisture divergence and temperature fields in the initial conditions for the two cases. Therefore the forecast of rainstorms is closer to observations in both quantity and pattern.