Transformer-based super-resolution downscaling for regional reanalysis: Full domain vs tiling approaches

This Zenodo repository provides the trained models and prediction outputs used in the study "Transformer-based super-resolution downscaling for regional reanalysis: Full domain vs tiling approaches". The study evaluates and compares the performance of several deep learning-based super-resolution (SR) methods, including Swin2SR, DeepESD, and U-Net, for enhancing the spatial resolution of reanalysis temperature data from ERA5 (0.25º) to CERRA (0.05º) over the Iberian Peninsula. 🔹 File: predictions_2019-2020.zip This archive contains super-resolved 2-meter temperature predictions (t2m) over the Iberian Peninsula for the years 2019 and 2020, generated by each of the downscaling methods evaluated in the paper. All files are in NetCDF (.nc) format, structured by model and downscaling strategy. Each file provides 3-hourly temperature fields at ~5.5 km resolution (CERRA grid, 0.05º). Structure: full-domain/ – Predictions from models trained over the entire domain: bicubic/predictions.nc: Bicubic interpolation baseline. deepesd/predictions.nc: DeepESD model predictions. swin2sr/predictions.nc: Swin2SR (Transformer) full-domain predictions. unet2d/predictions.nc: U-Net full-domain predictions. tiles/weighted-sampling/ – Predictions from models using domain tiling with weighted merging: swin2sr-p/predictions.nc: Swin2SR trained with patch-based sampling (40x40 overlapping tiles moving 20x20). swin2sr-t/predictions.nc: Swin2SR trained using a tiled layout (40x40 non-overlapping tiles). 🔹 File: models.zip This archive contains all trained model artifacts used to generate the predictions. It includes model weights (in PyTorch or ONNX format), configuration files, coordinate reference grids, and reproducibility metadata. Structure: full-domain/ – Models trained on the entire Iberian Peninsula: deepesd/, swin2sr/, unet2d/: model.onnx or model.pth: Trained model weights. config_data.yml, config_model.yml: Data and model configuration. features_coordinates.nc, labels_coordinates.nc: Input/output spatial grids. tiles/weighted-sampling/ – Tiling-based training strategies: swin2sr-p/ and swin2sr-t/: model.pth: Trained model weights in PyTorch. Full config files and coordinate NetCDFs as above. Additional files: highres_static_covariables.nc, lowres_static_covariables.nc: Auxiliary inputs (e.g., elevation, land cover) used in training. These models allow researchers to: Reproduce the prediction results in predictions_2019-2020.zip, Analyze different training strategies (full vs tiled), Fine-tune or reuse the models on new domains or variables, Compare architectures in a standardized setting.