The skill of statistical downscaling in future climate with high-resolution climate models as pseudo-reality

Abstract

Study region: Belgium Study focus: We assessed statistical downscaling (21 methods) skill and assumptions for seven extreme rainfall indicators in three Belgian areas. The ensemble consists of (i) perturbation methods with four members, (ii) quantile mapping with ten members, and (iii) machine learning with seven members. The methods were evaluated by a perfect predictor experiment in which the outputs of a high-resolution climate model are considered pseudo-observations. The ALARO-0 model with 4 km spatial resolution was considered the high-resolution climate model. The high-resolution outputs were compared with the downscaled results of three future scenarios of the 50 km resolution ALARO-0 model. New hydrological insights: The methods’ skill depends on the area. The skill is higher for the Belgian coastal area compared to the inland and hilly areas. Methods which are based on stationary assumptions have a good skill for the coastal area but show a lower skill for the other areas where the temporal variability in meteorological conditions is stronger. The higher this temporal variability, the more prone are the stationary assumptions to deteriorate in time under different climate forcing conditions. The quantile-based perturbation methods are overall the most robust. An ensemble approach is advised, where different downscaling methods are applied and the uncertainty in the downscaling taken into account, especially when a pseudo-future evaluation, as done in this study, is not possible.

Keywords

• Climate change
• Statistical downscaling
• Quantile mapping
• Machine learning
• Change factors
• Extreme precipitation