A case study of statistical downscaling in Australia using weather classification by recursive partitioning

Abstract

Because of their coarse resolution General Circulation Models (GCMs) lack the ability to predict reliably surface variables that may be needed for climate effects studies on a regional scale. Stochastic downscaling methods, based on the relationship between large-scale circulation types and regional surface variables, offer one approach to bridge this scale mismatch. The method explored in this paper uses recursive partitioning (also known as Classification Tree Analysis) to identify circulation patterns (weather states) in observed sea level pressure (SLP) fields that are most closely related to certain rainfall occurrence patterns (dry/wet) at multiple stations. The joint distribution functions of rainfall amounts at a set of stations are then used to define a statistical model for rainfall conditional on these weather states and the previous day's rainfall occurrence/absence. This model is applied to SLP fields from control, 2×CO 2 equilibrium and transient 100 year GCM integrations performed at the Max-Planck-Institut für Meteorologie, Hamburg to simulate local rainfall corresponding to the large-scale GCM climates. The weather generator is applied for several case studies in Australia. Four regions, one each in the West, the North, the Southeast and East of Australia are considered, for the summer and winter seasons, respectively. These regions correspond to different climate zones: the North has mainly summer monsoon rain, the West has dominant frontal winter rainfall, and moderate Southeast has a relatively uniform rainfall distribution throughout the year. Thus, the performance of the weather generator in different climatic regimes can be evaluated. Since rainfall amounts are simulated by sampling from historical data a statistical uncertainty is attached to the generated rainfall sequences which is due to the procedure used. This is reflected in uncertainties of the downscaled mean rainfall of generally less than ±10%, with larger numerical intervals for dry stations and seasons. The effects of a 2×CO 2 signal in the equilibrium run on local rainfall are not dramatic. For the transient experiment, there is an indication of more rainy days and a decrease in the length of dry periods in West Australia in winter. The strongest change is identified in the North in the monsoon season which would have decreasing rainfall. No unequivocal changes were found for the Southeast. However, consistent with earlier studies, the bias of GCM control simulations strongly affects the results, as the difference between precipitation statistics based on observed and control run pressure fields often exceeds the difference between GCM control and altered climate. Considering this deficiency in current GCM simulations, the disagreement between different climate models and the uncertainty attached to the statistical downscaling method, not too much reliance can be put on the relevance of these results for regional hydrologic impact modeling.