Comparison of future runoff projections using different downscaling methods.

Global warming leads to changes in future rainfall and runoff that has significant impact on the regional hydrology and water availability. All the large-scale climate impact studies use the future climate projections from global climate models (GCMs) to estimate the impact on future water availability. Downscaling techniques are commonly used in modelling climate impact on hydrology studies to transform the lower resolution climate model outputs to a finer scale suitable for regional or catchment scale impact assessment. There are a number of downscaling approaches reported in literature ranging from simple to complex. Many impact studies rely on an empirical scaling method (also referred to as ‘change factors (CFs)’, ‘perturbation method’ or ‘delta-change method’ in various studies) that scales the historical record based on GCM outputs. More complex methods include statistical downscaling that establish relationships between large scale predictors and local scale predictands and computationally intensive dynamical downscaling that embeds a higher resolution regional climate model (RCM) within a GCM.
In light of the diversity of the available downscaling approaches, this study aimed to explore the merits and limitations of different downscaling methods in modelling climate change impact on runoff. To do so, we first assessed the historical daily runoff and salient runoff characteristics modelled using rainfall from three widely used downscaling methods – Analogue, NHMM, and WRF – and compared them with the modelled runoff using observed rainfall. We then compared the future runoff projections (change in future runoff) derived from these downscaling methods. The modelling experiments were carried out for eight unregulated catchments in south-eastern Australia. The downscaling models were driven by NCEP/NCAR reanalysis data (NNR) and one GCM (CSIRO mk3.5). Two hydrological models – GR4J and Sacramento –were calibrated against observed streamflow data and used to model historical and future runoff. The SILO Data Drill was used for validation.
The statistical downscaling methods (Analogue and NHMM) performed well in reproducing historical rainfall and runoff characteristics when driven by NNR. But this was not necessarily the case when they were driven by historical GCM simulations. The performance of the dynamical downscaling method (WRF) was sensitive to the spatial resolution it was applied and similar amount of biases are observed when driven by both NNR and GCM. The systematic biases in the downscaled rainfall highlight the need for applying a bias correction method. However, our experiments showed that a simple linear bias correction method need to be used with caution as it can adjust the rainfall distribution in a wrong way which lead to larger biases in runoff. Furthermore, the benefit of the linear bias correction method on future runoff projections is unclear as it does not seem to have major impact on change in runoff characteristics. The results indicate that there can be substantial differences between the future runoff projections from different downscaling methods. The differences for mean annual runoff between the different downscaling methods are usually smaller compared to those for other salient runoff characteristics.