Continental-scale bias correction and random forest downscaling of CMIP6 precipitation across Europe

This study evaluated bias-correction-based downscaling approaches for precipitation data from 11 high-resolution CMIP6 models across Europe. The empirical quantile mapping technique demonstrated superior performance by aligning the model outputs with observed precipitation data. Random forest-empirical quantile mapping model outperformed the random forest model, particularly for low rainfall intensities (0-5 mm/day), and exhibited residual errors closer to zero across most European regions. Precipitation overestimations by CMIP6 models in Central and Eastern Europe were significantly reduced through the application of empirical quantile mapping, random forest and random forest empirical quantile mapping approaches. Seasonal analysis revealed lower residual errors in summer than in winter for the evaluated methods. Future projections will provide insights into the reduced overestimation of high quantiles, leading to reliable precipitation estimates. The proposed methodological framework suggests that integrating bias correction complemented by machine learning techniques enhances the accuracy of regional precipitation downscaling, which is crucial for climate risk management in Europe.