A new methodology for downscaling seasonal climate data to daily rainfall at a collection of gauging sites in the framework of Dynamic Bayesian Networks is presented. The method extends the existing nonhomogeneous hidden Markov models (NHMM) in a number of ways. First, explicit dependence on prognostic variables instead of hidden states is considered. This allows the direct consideration of physically meaningful variables derived from GCM simulations or observed pre-season climatic indices as conditioning variables for daily rainfall occurrence and amount probability distributions. Second, the ability to consider continuous dependence of rainfall states and their daily transition probabilities addresses the curse of dimensionality that impacts NHMM as the number of hidden states increases. Third, a hierarchical Bayesian approach is used to model the spatial downscaling process, including the macro scale and local scale spatial dependence in the model parameters. Thus, the common regional effect of the predictors and the separate station effects can be identified and modeled jointly. Fourth, model parameters are allowed to be temporally variable recognizing that model “skill” may vary from year to year depending on the underlying climate state. The temporal variation is modeled as a random walk whose variance is inferred from the data. Applications of the methodology to data from the Everglades National Park region in S. Florida are presented. Comparisons with a NHMM model are also provided. The predictors used are ensemble means of seasonal rainfall and OLR forecasts from the DEMETER and ECHAM models. Trends from the spatial and temporal variation of model parameters are identified and interpretations are provided in terms of their relationship to slowly varying SST conditions – predominantly the tropical Atlantic seems to be well related to the Florida rainfall. Thus, biases in the GCM forecasts can be identified, related to observed SST fields whose response is not captured, and addressed by including those variables as predictors.