SR-Weather bridges the resolution gap between global machine-learning weather prediction and kilometer-scale local temperature by integrating high-resolution terrain, urban-surface, and climatological information into a forecast-oriented super-resolution framework.
Read Paper ↗Global machine-learning weather models provide accurate and computationally efficient forecasts, but their coarse spatial resolution limits the representation of terrain-driven temperature gradients, urban heat islands, and other localized surface effects. SR-Weather addresses this scale gap through learned super-resolution.
The novelty of SR-Weather lies not only in increasing spatial resolution, but in connecting machine-learning weather prediction with local-scale environmental information and demonstrating that super-resolution can also improve forecast accuracy.
Rather than treating downscaling only as a reanalysis reconstruction problem, SR-Weather is designed for direct application to machine-learning weather forecasts. A model trained with ERA5 is transferred to FuXi medium-range forecasts without retraining.
Forecast TransferFine-scale terrain, impervious-surface, and seasonal climatological information explicitly condition the reconstruction so that local temperature variability is recovered beyond what is available in the coarse meteorological grid alone.
Physical ContextThe framework improves spatial detail while also reducing forecast error. This shows that learned super-resolution can operate as more than interpolation: it can refine the statistical quality of machine-learning weather predictions.
Forecast RefinementThe model is first trained using ERA5 2-m air temperature as the coarse input and MODIS-derived 1-km air temperature as the target. The trained model is then applied directly to FuXi forecasts to generate kilometer-scale temperature forecasts.
Daily 1-km air temperature is derived from MODIS land-surface temperature using station observations and environmental predictors.
SR-Weather learns the mapping from 0.25° ERA5 temperature to MODIS-derived 1-km air temperature while integrating fine-scale spatial predictors.
The ERA5-trained model receives FuXi 0.25° forecasts and produces 1-km temperature fields for lead times from one to seven days.
SR-Weather extends SE-SRCNN by combining global average pooling with global maximum and minimum pooling. DEM and the seasonal climatology map guide the extrema-sensitive attention branches, helping the model preserve localized thermal contrasts.
In the ERA5 test experiment, SR-Weather achieved the strongest overall RMSE and R² among the evaluated methods. Improvements were observed across high-elevation, low-elevation, and urban regions.
Relative to bicubic interpolation, SR-Weather reduced RMSE by 46.4% in high-elevation areas, 30.6% in low-elevation areas, and 21.6% in urban areas.
Relative to bicubic interpolation in the terrain-stratified evaluation.
The ERA5-trained network is applied to FuXi forecasts without additional model retraining. SR-Weather improves agreement with station observations while reconstructing urban and terrain-induced temperature structures.
The paper reports that the downscaled FuXi 7-day forecast achieved lower error than a 1-day low-resolution forecast processed with bicubic interpolation, indicating bias correction in addition to resolution enhancement.
Approximate reduction in computational expense per grid cell compared with the LDAPS-level numerical forecast estimate.
