Optimizing and verifying an ensemble-based rainfall model

Abstract

In this thesis, I modified, optimized, and verified the stochastic Recursive Cluster-point Rainfall model of Chatdarong (2006). A novel error metric allows comparison of the stochastic ensemble of rainfall image forecasts to a single observation (radar) image. The error metric forgives position errors and provides a flexible framework for assessing how well the model works vis-a-vis a set of image measures, including the distribution of rainfall intensities over the domain at different scales. The error metric is used in various forms to perform ad hoc optimization of the model parameters and to verify the ensemble forecast in a probabilistic framework. Verification results show that the optimized model is limited in its ability to create truly realistic rainfall patterns. Despite the model's limitations, it has unique applicability to ensemble-based rainfall data assimilation.