A Two-Stage Bias-Correction and Super-Resolution Framework for Post-Processing Climate Model Outputs

Abstract. General circulation models (GCMs) underpin climate change assessments, yet their coarse spatial resolution and systematic biases constrain their direct use in regional applications. Post-processing approaches such as bias correction and statistical downscaling are therefore widely applied, yet these steps are often implemented independently, leading to inconsistencies between corrected statistics and spatial structure. This study presents a reproducible two-stage framework that integrates Quantile Delta Mapping (QDM) for bias correction with a deep learning-based super-resolution method to improve the statistical fidelity and spatial detail of climate model outputs. The framework is evaluated using precipitation, runoff, and evapotranspiration from three Coupled Model Intercomparison Project Phase 6 (CMIP6) GCMs (Canadian Earth System Model (CanESM5), Hadley Centre Global Environment Model (HadGEM3‐GC31‐LL), and Max Planck Institute Earth System Model (MPI‐ESM1‐2‐HR)), over the Hadejia-Jama'are River Basin in northern Nigeria. We first demonstrate that raw model outputs exhibit substantial biases, with domain-averaged root mean square errors (RMSE) of 48.6–57.3 mm month⁻¹ for precipitation, 0.93–7.51 mm month⁻¹ for runoff, and 33.8–58.7 mm month⁻¹ for evapotranspiration. QDM substantially reduces systematic errors, lowering precipitation RMSE to 23.8–27.8 mm month⁻¹, runoff RMSE to 0.24–1.85 mm month⁻¹, and evapotranspiration RMSE to 3.7–4.3 mm month⁻¹, while preserving projected distributional changes, as confirmed by Kolmogorov-Smirnov (D ≤ 0.072) and Wasserstein (≤ 11.86) metrics. In the second stage, a conditional Generative Adversarial Network (GAN) super-resolves the bias-corrected precipitation fields from ~250 km to 14 km (regridded to 1 km), outperforming bilinear and bicubic interpolation in terms of structural similarity and spatial coherence. The proposed QDM-GAN framework is fully documented and reproducible, with openly available code and data sources, and is intended as a modular post-processing tool that can support downstream modeling applications requiring bias-corrected, high-resolution climate inputs.