A simple hybrid statistical-dynamical downscaling method for emulating regional climate models over Western Europe. Evaluation, application, and role of added value?

A hybrid statistical dynamical downscaling method intended to emulate regional climate models is described and applied to Western Europe. The method is based on a constructed analogues algorithm, already used for statistical downscaling. For emulation, the statistical downscaling relationship is not derived from observations but from climate projections at low and high resolution. The hybrid approach therefore does not rely on the stationarity assumption inherent to conventional statistical downscaling. Within a perfect model framework, and using a large number of regional projections, the hybrid method is shown to reproduce climate change signals very well and to outperform a conventional statistical downscaling method also based on constructed analogues. The hybrid approach remains skillful even when applied to very low resolution climate data. In practice, two emulation modes exist. In the GCM/RCM mode, the downscaling relationship is built between a RCM and its forcing GCM. In the RCM/RCM mode, the relationship is built between a RCM and the same RCM after aggregation of its results to a low resolution grid. The large-scale climate change signal of the downscaled GCM is generally retained with the RCM/RCM mode, but not with the GCM /RCM mode. Additionally, the choice of the GCM/RCM pair used for learning leads to large differences in downscaling results at large scale (i.e. at low resolution) with the GCM /RCM mode, but not with the RCM/RCM mode. These results are explained by the differences that generally exist at large scale between projected changes by current RCMs and their forcing GCMs. Whether these differences are a testimony of a real added value of RCMs at large scale in the climate change context, or whether they have other causes, is therefore a crucial question.