Spatially generalizable bias correction of satellite solar radiation for regional climate assessment—a case study in Japan

Evaluations of model- and reanalysis-based surface solar radiation (SSR) products often depend on satellite-derived estimates, which can be biased, particularly under broken-cloud conditions or over high-albedo surfaces. In this study, we first evaluated JAXA’s Himawari satellite SSR against ground-based observations. We then combined sparse but accurate ground measurements with high-resolution meteorological data to train an eXtreme Gradient Boosting (XGBoost) model that captures and spatiotemporally extrapolates the systematic bias between satellite SSR and ground truth data. Our model is physics-informed, interpretable via feature-attribution analyses, and generalizes across snow-covered terrain. Applying this correction over Japan, we achieved close spatial pattern alignment with a distinct ground-based gridded SSR dataset, which further allowed us to examine the spatial transferability of the correction method. Finally, we used the bias-corrected satellite SSR data to assess a novel high-resolution regional reanalysis, and found that the reanalysis data reproduce expected spatial patterns and overall SSR magnitudes well but tend to slightly overestimate the SSR data under cloudy skies. However, especially under clear skies, the reanalysis remains valuable for refining bias-corrected satellite SSR data in certain high-altitude mountainous regions, where our machine learning model struggles to extrapolate. Overall, the proposed method shows promise for application across platforms and regions, enabling more reliable datasets for climate assessment and downscaling.