Data and trained models — Deep-learning statistical downscaling of daily precipitation over West Africa

Supporting data and trained model weights for the study on deep-learning statistical downscaling of daily precipitation over West Africa (ERA5 0.25° → IMERG 0.1°, factor 2.5×). Files: arrays_01deg.npz — cached model outputs and IMERG observations on the 0.1° evaluation grid (7 configurations + observations), allowing reproduction of all figures and metric tables without re-running the networks. weights.zip — the six trained PyTorch models (U-Net Simple/SE/CBAM and their cGAN-fine-tuned variants), seed 42, as reported in the body of the article. weights_appendices.zip — the eighteen models of the appendices: the three U-Net architectures at seeds 123 and 456, DeepSD in its three configurations at three seeds, and U-Net-CBAM trained with the ERA5 precipitation channel at three seeds, with the matching normalisation statistics. arrays_01deg_s42.npz, arrays_01deg_s123.npz, arrays_01deg_s456.npz — predicted fields of the three architectures at each seed, on the 0.1° grid. The replication code is available on GitHub: github.com/sagna1/downscaling-westafrica. Raw predictors (ERA5) and target (IMERG Final Run V06) are openly available from the C3S Climate Data Store and NASA GES DISC respectively.