This study examines the ability of a regional spectral model to assimilate high resolution rainfall data over South America using a physical initialization procedure. Physical initialization was developed at Florida State University (FSU) to improve rainfall prediction in numerical weather prediction models. By assimilating observed rainfall rates using reverse cumulus algorithms and nudging, the model reproduces the proper rainfall rates at the end of the assimilation period (typically 24 hours) and subsequently the forecast precipitation is improved. Experiments using the recently developed FSU Nested Regional Spectral Model (FSUNRSM) were made with and without rainfall initialization. The convective-stratiform technique (CST) was applied to estimate precipitation fields from infrared Geostationary Operational Environmental Satellite (GOES-IR) data. This technique looks for minima in the brightness temperature field to define thunderstorms locations. Comparisons between CST using GOES-IR images and FSUNRSM accumulated precipitation fields indicate a large improvement for the initial condition fields using such rainfall assimilation procedure.