A Two-Stage Downscaling Approach using Machine Learning and image super-resolution techniques for high-resolution seasonal climate forecasts in the Alpine region

The Alpine region faces heightened risks from climate change due to its complex terrain and ecosystems, highlighting the significant global challenge posed by a warming climate. The region is particularly susceptible to the effects of global warming, which not only intensifies weather extremes but also significantly impacts hydrological processes. These changes increase the frequency and severity of extreme events like droughts and floods, further heightening the region's vulnerability. Accurate local climate predictions are essential for effectively managing these risks, as they provide the spatial and temporal precision necessary for hydrological simulations. Such high-resolution data enable detailed modelling of water availability, runoff patterns, and flood risks, facilitating improved planning and adaptation strategies. However, existing global datasets often lack the resolution needed for these assessments. To address this gap, this research aims to generate high-resolution seasonal climate forecasts specifically designed for the Alpine region, providing an essential tool for understanding climate variability, managing hazards, and supporting hydrological analyses. The study proposes a novel two-stage downscaling approach within the perfect prognosis framework to enhance the spatial resolution of ECMWF (European Centre for Medium Range Weather Forecasts) SEAS5 (Seasonal Forecast System 5) seasonal forecasts from native 0.25°x0.25° to 1 km for the Alpine region. Key variables include daily temperature, precipitation, and downward surface solar radiation. In the first stage, pixel-by-pixel downscaling is performed though LGBM (Light Gradient Boosting Machine) regression applied to ERA5 reanalysis predictor fields matched against CHELSA-W5E5 (v1.1) fields, conservatively interpolated to 6-km resolution. Predictors are selected through feature importance analysis via cluster-based regression and is optimized for the 2005–2016 training period. The trained model is then applied to the 51 ensemble members of SEAS5 predictors, generating target variables at a 6 km resolution. In the second stage, the 6-km downscaled outputs, along with additional static predictors such as elevation, aspect, and cyclically encoded day of the year, are passed to a sliding-window Enhanced Super-Resolution Generative Adversarial Network (ESRGAN). This image super-resolution technique trained and optimized using CHELSA-W5E5 at its native 1-km resolution, further refines the forecasts to produce high-resolution seasonal predictions with 51 ensemble members at 1 km resolution. The two-stage scheme was found to improve the downscaling performance with respect to the application of one-step method. The contribution will present the overall methodology and the results of the model evaluation. The outcomes of this study are expected to play a key role as critical inputs for a drought prediction module within the framework of the EU-funded interTwin project. This research has been funded by the European Union through the interTwin project (101058386).