Through numerical model sensitivity experiments, the impacts of land-surface process parameterization on the model predictability of two heavy rainfall events are studied. The experiments of closing land-air fluxes in the model show that for advective precipitation, such as the severe flood near the Huaihe River basin during 3－4 July 2003, the land-air fluxes contribute little and the model forecast represents small dispersion or relatively high predictability. For evaporative rainfall, such as the shower during 2－3 August 2003, the land-air fluxes contribute much and the model forecast represents large dispersion or relatively low predictability. The experiments of replacing different land processes in the model show that the error growth exhibits selectivity, which only grows rapidly and quickly in the precipitation area and large wind speed (or jet) areas. The errors and instability energy should come from the same origin, and they increase simultaneously. That means model predictability and uncertainty are interlaced. In addition, if the model starts to run at night other than in the daytime, the errors should develop slowly and the predictability time could be prolonged.