Stochastic downscaling of numerical climate model simulations

Interest in stochastic downscaling has grown from the inability of general circulation models (GCMs) and limited area models (LAMs) to reproduce observed daily precipitation statistics at local and regional scales under present climate conditons. Although GCMs and LAMs perform reasonably well in simulating synoptic atmospheric fields, they tend to over-estimate the frequency and under-estimate the intensity of daily precipitation. This paper describes the application of a nonhomogeneous hidden Markov model (NHMM) fitted to observed atmospheric and precipitation data to a LAM simulation for South–West Western Australia. The NHMM is unique amongst stochastic downscaling methods in that it determines the most distinct patterns in a multi-site, precipitation occurrence record rather than patterns in atmospheric fields. These patterns can in turn be defined as conditionally dependent on a range of daily atmospheric series. We compare the LAM simulation and the downscaled LAM simulation with observed ‘winterʹ precipitation statistics at six stations near Perth, Western Australia. The results show that the downscaled simulations reproduce observed precipitation probabilities and wet and dry spell frequencies at each station. The LAM simulation tends to under-estimate the frequency of dry spells and over-estimate the probability of precipitation and the frequency of wet spells.