Realistic rainfall ensembles are crucial for hydrologic applications of ensemble predictions and data assimilation. In particular, ensemble prediction and assimilation algorithms can be expected to work better when the random precipitation events used to derive sample covariances or to force the land surface models have a spatio-temporal structure similar to observed rainfall. In this paper we present a procedure for generating rainfall replicates from on hourly GOES cloud top temperature data and space-time statistics derived from NOWRAD radar data. The first step in this procedure uses the GOES data to determine the size and shape of rainfall clusters. The second step simulates multiple realizations of rainfall within these clusters. The number of rainfall peaks (local maxima of rain intensity) within a particular cluster is determined from the cluster size. The locations and intensities of these peaks are generated randomly using kernel functions derived from NOWRAD image statistics. The algorithm's ability to generate replicates that are statistically similar to observations is verified using a multiattribute rank histogram analysis. The method is illustrated with an example that covers the United States Great Plains (at 0.5 degree spatial resolution) during the summer of 2004.