Statistical downscaling of daily rainfall for hydrological impact assessment

Abstract

Water resource management and planning increasingly need to incorporate the effects of global climate change on regional climate variability in order to accurately assess future water supplies. Therefore future climate projections, particularly of rainfall, are of utmost interest to water resource management and water-users. General circulation models (GCMs) are the primary tool used to simulate present climate and project future climate. The outputs of GCMs are useful in understanding how future global climate responds to prescribed greenhouse gases emission scenarios. However GCMs do not provide realistic daily rainfall at scales below about 200 km, at which hydrological processes are typically assessed. Statistical downscaling techniques have been developed to resolve the scale discrepancy between GCM climate change scenarios and the resolution required for hydrological impact assessment, based on the assumption that large-scale atmospheric conditions have a strong influence on local-scale weather. Gridded rainfall is important for a variety of scientific and engineering applications, including climate change detection, the evaluation of climate models, the parameterization of stochastic weather generators, as well as assessment of climate change impacts on regional hydrological regimes and water availability, whereas statistical downscaling has predominantly provided daily rainfall series at the site (point) scale. This study explores the application of statistical downscaling to gridded and catchment rainfall datasets using three methods: 1) statistically downscaling to sites and then post-processing to interpolate to gridded rainfall; 2) treating each grid cell as an observed site for statistical downscaling directly; and 3) treating each sub- catchment as an observed site and statistically downscaling to sub-catchment averaged rainfall. The statistical downscaling Nonhomogeneous Hidden Markov Model (NHMM), which models multi-site patterns of daily rainfall as a finite number of 'hidden' (i.e. unobserved) weather states, is used for a study region comprising several catchments of the southern Murray-Darling Basin (MDB) in south-eastern Australia, which until this year has been experiencing a decade long drought. The results show that: 1) the best performance, of the methods compared, resulted from calibration to meteorological station data. The NHMM calibrated to 38 stations across the lower MDB reasonably reproduced the validation period mean rainfall characteristics; 2) Calibration to catchment average rainfall, for a corresponding set of 38 sub-catchments, produced a reasonable calibration result but significantly more bias for the validation period in comparison with the station NHMM results. Whether the threshold used to define wet-days influences this aspect of performance will be investigated in future work; 3) Calibration to all 364 grid cells across the study area catchments produced a biased result for both calibration and validation periods. Given this, subsequent calibration to grid cells for 12 smaller sub-catchments produced mixed results given numerical instabilities in the NHMM optimization algorithms; and 4) It is difficult to determine the relative contribution to validation-period bias that could be the result of several factors, such as inadequate parameterizations of the NHMMs, non-stationary in the relationship between NNR predictors and rainfall data, or data quality limitations. Most probably all are involved to some extent, and so future work should also investigate validation issues.