An Evaluation of Statistical and Dynamical Techniques for Downscaling Local Climate

An assessment is made of downscaling estimates of screen temperature and precipitation observed at 976 European stations during 1983–94. A statistical downscaling technique, in which local values are inferred from observed atmospheric predictor variables, is compared against two dynamical downscaling techniques, based on the use of the screen temperature or precipitation simulated at the nearest grid point in integrations of two climate models. In one integration a global general circulation model (GCM) is constrained to reproduce the observed atmospheric circulation over the period of interest, while the second involves a high-resolution regional climate model (RCM) nested inside the GCM. The dynamical and statistical methods are compared in terms of the correlation between the estimated and observed time series of monthly anomalies. For estimates of temperature a high degree of skill is found, especially over western, central, and northern Europe; for precipitation skill is lower (average correlations ranging from 0.4 in summer to 0.7 in winter). Overall, the dynamical and statistical methods show similar levels of skill, although the statistical method is better for summertime estimates of temperature while the dynamical methods give slightly better estimates of wintertime precipitation. In general, therefore, the skill with which present-day surface climate anomalies can be derived from atmospheric observations is not improved by using the sophisticated calculations of subgrid-scale processes made in climate models rather than simple empirical relationships. It does not necessarily follow that statistical and dynamical downscaling estimates of changes in surface climate will also possess equal skill. By the above measure the two dynamical techniques possess approximately equal skill; however, they are also compared by assessing errors in the mean and variance of monthly values and errors in the simulated distributions of daily values. Such errors arise from systematic biases in the models plus the effect of unresolved local forcings. For precipitation the results show that the RCM offers clear benefits relative to the GCM: the simulated variability of both daily and monthly values, although lower than observed, is much more realistic than in the GCM because the finer grid reduces the amount of spatial smoothing implicit in the use of grid-box variables. The climatological means are also simulated better in the winter half of the year because the RCM captures some of the mesoscale detail present in observed distributions. The temperature fields contain a mesoscale orographic signal that is skillfully reproduced by the RCM; however, this is not a source of increased skill relative to the GCM since elevation biases can be largely removed using simple empirical corrections based on spatially averaged lapse rates. Nevertheless, the average skill of downscaled climatological mean temperature values is higher in the RCM in nearly all months. The additional skill arises from better resolution of local physiographical features, especially coastlines, and also from the dynamical effects of higher resolution, which generally act to reduce the large-scale systematic biases in the simulated values. Both models tend to overestimate the variability of both daily and monthly mean temperature. On average the RCM is more skillful in winter but less skillful in summer, due to excessive drying of the soil over central and southern Europe. The downscaling scores for monthly means are compared against scores obtained by using a predictor variable consisting of observations from the nearest station to the predictand station. In general the downscaling scores are significantly worse than those obtained from adjacent stations, indicating that there remains considerable scope for refining the techniques in future. In the case of dynamical downscaling progress can be made by reducing systematic errors through improvements in the representation of physical processes and increased resolution; the prospects for improving statistical downscaling will depend on the availability of the observational data needed to provide longer calibration time series and/or a wider range of predictor variables.