WeatherAdaptSR

Transferring Knowledge Across Regions: Domain Adaptation for Super-Resolution Km-Scale Models

Downscaling improves weather and climate predictions by enhancing spatial and temporal resolution, supporting decision-making, risk modelling, and resource management. Super-resolution (SR) models, which downscale low-resolution data to high-resolution outputs for variables like temperature and precipitation, have proven effective when trained on regional data. However, their ability to generalize across diverse climates and topographies remains uncertain. This study investigates domain adaptation techniques, such as feature and latent space alignment, to enhance SR model transferability. Using a 10-year COSMO simulation over Europe, we coarsen data to 18km resolution before downscaling it back to 2.2km, assessing performance across different landscapes. We show the potential of domain adaptation to improve model generalization.