The influence of the soil moisture initialization on predicting rainfall in a limited-area model (LAM) was analysed. The model selected for the simulation was the LAM of Bologna (LAMBO). The case study selected in this paper was the flash flood episode that occurred during September 1995 in the South Ticino area (Switzerland). Two different experiments were conducted. In the first experiment, all initial fields including the initial soil moisture were provided by European Centre for Medium-range Weather Forecast (ECMWF) analyses. In the second one, the initial soil moisture fields were calculated by running a soil–vegetation–atmosphere transfer scheme (the Land Surface Process Model—LSPM) in stand-alone mode. The LSPM was driven by the synoptic observations carried out on a mesoscale area that includes Northern Italy. In order to overcome the well-known problem of the initial soil moisture content, LSPM was run for a long period (5 months) preceding the flood, in order to reach a stability state. The soil moisture field calculated by LSPM is more structured than the ECMWF one, and the numerical values of soil moisture are generally lower than those of ECMWF. The results seemed to indicate that the precipitation field obtained for the flood episode in the second experiment is closer to the observations than in the control run, even if there were still some underestimations. The precipitation peak modelled on South Ticino was closer to the observations, and this result seems to confirm the importance of the correct initialization of the soil moisture field even in cases of short-range forecasting. Copyright © 2002 John Wiley & Sons, Ltd.