A nonparametric stochastic downscaling framework for daily rainfall at multiple locations

Use of General Circulation Models (GCMs) for climate change impact assessment is often limited by their incapability at representing local features and dynamics at spatial scales finer than the in-built GCM grid scale. This has led to the development of downscaling techniques for transfer of coarse GCM simulated weather output to finer spatial resolutions. This paper presents a nonparametric stochastic spatial downscaling framework for multisite daily rainfall occurrence and amount. At site rainfall occurrences are downscaled using a nonparametric nonhomogeneous hidden Markov model (NNHMM) that represents spatial dependence across the rainfall occurrence field using a dynamic weather state indicative of the centroid and average wetness fraction of the rainfall occurrence field. The rainfall amounts on the wet days are downscaled using a nonparametric kernel density approach that accommodates variations in the rainfall downscaling model at individual locations. Spatial dependence in the rainfall amounts is simulated by driving each of the single-site amounts model with spatially correlated random numbers. The proposed framework is applied for downscaling of rainfall at a network of 30 rain gauge stations around Sydney in Australia, and its performance is evaluated. The analyses of the results show that the logic of providing separate treatments for rainfall occurrence and amounts at individual locations imparts considerable accuracy in the representation of characteristics of interest in hydrologic studies. These characteristics include representation of rainfall spell patterns, spatial distribution of the rainfall occurrence and amount fields, representation of low and high rainfall extremes at individual stations and across the field, as well as common indicators of water balance and variability that are of importance in a catchment scale water balance simulation.