Non-linear statistical downscaling of present and LGM precipitation and temperatures over Europe

Local-scale climate information is increasingly needed for the study of past, present and future climate changes. In this study we develop a non-linear statistical downscaling method to generate local temperatures and precipitation values from large-scale variables of a Earth System Model of Intermediate Complexity (here CLIMBER). Our statistical downscaling scheme is based on the concept of Generalized Additive Models (GAMs), capturing non-linearities via non-parametric techniques. Our GAMs are calibrated on the present Western Europe climate. For this region, annual GAMs (i.e. models based on 12 monthly values per location) are fitted by combining two types of large-scale explanatory variables: geographical (e.g. topographical information) and physical (i.e. entirely simulated by the CLIMBER model). <br><br> To evaluate the adequacy of the non-linear transfer functions fitted on the present Western European climate, they are applied to different spatial and temporal large-scale conditions. Local projections for present North America and Northern Europe climates are obtained and compared to local observations. This partially addresses the issue of spatial robustness of our transfer functions by answering the question "does our statistical model remain valid when applied to large-scale climate conditions from a region different from the one used for calibration?". To asses their temporal performances, local projections for the Last Glacial Maximum period are derived and compared to local reconstructions and General Circulation Model outputs. <br><br> Our downscaling methodology performs adequately for the Western Europe climate. Concerning the spatial and temporal evaluations, it does not behave as well for Northern America and Northern Europe climates because the calibration domain may be too different from the targeted regions. The physical explanatory variables alone are not capable of downscaling realistic values. However, the inclusion of geographical-type variables – such as altitude, advective continentality and moutains effect on wind (W–slope) – as GAM explanatory variables clearly improves our local projections.