AINPP PB LATAM DATASET

This dataset provides a curated and harmonized collection of gridded precipitation fields derived from the GSMaP Near Real-Time (NRT) and GSMaP Moving Vector with Kalman filter (MVK, V8) products, specifically designed to support reproducible machine learning experiments in precipitation nowcasting. The dataset is structured to enable supervised learning, where GSMaP-NRT fields are used as input data and the higher-quality GSMaP-MVK fields serve as forecast targets, allowing systematic evaluation of models under realistic data latency constraints. The dataset covers the AINPP Latin America domain (latitudes −55° to 33°, longitudes −120° to −23°) on a regular 0.1° grid, yielding spatial fields of 880×970 pixels per time step. The temporal extent spans January 1, 2018, to December 31, 2024, with an hourly temporal resolution. To prevent temporal data leakage and ensure fair model development, the data are explicitly partitioned into training (2018–2022), validation (2023), and test (2024) subsets. Both products go through the same preprocessing pipeline, which includes a log1p transformation and then Z-score normalization. The normalization parameters are computed exclusively from the GSMaP-MVK training period and consistently applied to both NRT and MVK across all temporal splits. The dataset is saved in a compressed Zarr (v2) format that is organized for deep learning, and the normalization parameters are shared so that we can easily convert the data back to actual precipitation measurements. This dataset provides a robust and scalable foundation for precipitation nowcasting benchmarks and reproducible experimentation.