Downscaling daily climate data with super-resolution and CMIP6 global climate models to predict the climate in China

Traditional downscaling methods are often hampered by high computational costs and assumptions of statistical stationarity, which compromise the reliability of future regional climate risk assessments. To address these problems, we developed a downscaling framework that integrates super-resolution (SR) deep learning from computer vision into climate science by systematically comparing and evaluating eight distinct SR architectures. The optimal model was identified through a comparative evaluation of these super-resolution models, and the GCM climate data from CMCC-ESM2 and NorESM2-MM were downscaled to 10 km across China. The residual channel attention network (RCAN) model outperformed the other models in downscaling the four selected climate variables (tas, pr, rsds and sfcWind). The RCAN model achieved the highest R² values for tas (0.996), pr (0.906), rsds (0.992) and sfcWind (0.965), along with the lowest mean absolute errors (MAEs) for tas (0.428 K), pr (0.540 mm), rsds (3.329 W/m²) and sfcWind (0.172 m/s). In addition, the downscaling performance achieved with RCAN had minimal errors and stable results at annual, seasonal and monthly scales. Seasonal R² values for tas consistently exceeded 0.994, and the monthly MAE for sfcWind showed a minimal fluctuation amplitude of only 0.025 m/s throughout the year. Seasonal fluctuations were noted in rsds and pr during summer, whereas tas and sfcWind errors remained stable throughout the year. Future climate projections for 2015–2100 based on downscaled GCMs show marked warming trends, with temperature increases of up to 0.83 K per decade under SSP-585 and precipitation increasing by up to 23.46 mm per decade. Solar radiation trends vary by scenario and region, whereas wind speeds decline consistently across all scenarios. This framework, together with the optimally selected RCAN model, demonstrates a strong capability to generate high-fidelity 10-km climate data that effectively capture local topographic and climatic features. Such an enhancement substantially improves the precision of regional climate change assessments, providing a more robust scientific foundation for developing targeted adaptation strategies and supporting informed decision-making.