Temporally and spatially variable rainfall replicates are frequently required for hydrologic ensemble forecasting and data assimilation algorithms. Such algorithms can be expected to work better when their rainfall replicates more closely resemble observed storms. In this paper we present a new probabilistic procedure for generating realistic rainfall replicates. The procedure generates clusters of non-zero rainfall intensity within regions of remotely sensed GOES low cloud top temperature. The cluster boundaries are obtained from a multipoint geostatistical algorithm that uses NOWRAD training images to derive a random spatial support for each replicate. A truncated multiscale tree is used to generate rain rates within each cluster. The required tree parameters are estimated from the NOWRAD training images. A computational experiment based on Summer 2004 data from the Central US region indicates that the rainfall replicates simulated by the procedure are visually and statistically similar to individual NOWRAD images observed for the same storms and to a large ensemble of NOWRAD images collected throughout the summer simulation period.