Evaluation of statistical downscaling methods for climate change projections over Spain: Present conditions with imperfect predictors (global climate model experiment)

Statistical downscaling of climate projections is an active field of research and numerous intercomparison exercises of different techniques exist. Most evaluation studies make use of perfect predictors from a reanalysis, but neglect the analysis of how imperfect predictors from global climate models (GCMs) affect statistical downscaling. In this paper we evaluate and intercompare five statistical downscaling methods: (a) Analog, (b) Regression, (c) Artificial Neural Networks, (d) Support Vector Machines and (e) Kernel Ridge Regression, over a large ensemble of 18 GCMs participants in the CMIP5 project. These methods have been used to downscale maximum/minimum daily temperatures and daily accumulated precipitation on a high-resolution observational grid (0.05 degrees) over mainland Spain and the Balearic Islands. Our analysis focuses on the marginal aspects (mean values and tails) of the distributions by season. Results when driven by reanalysis and by GCMs have been compared in order to analyse the sensitivity of statistical downscaling methods to the use of imperfect predictors. Both for temperature and for precipitation all methods achieve, as expected, better results when applied over reanalysis, although no method displays a clearly higher sensitivity to the use of imperfect predictors from GCMs. We have found the existence of outlier GCMs, showing significantly higher errors than the rest of the ensemble. And we have quantified errors over GCMs, so this experiment provides a more precise idea of the reliability and accuracy of the overall performance (i.e., simulations by GCMs plus statistical downscaling) than the perfect predictor experiment and reveals an important source of uncertainty coming from GCMs even in a historical period.