Google DeepMind Opens Two Weather Datasets for AI Forecasting

Google DeepMind has opened two weather datasets for researchers, pairing hourly global satellite imagery with 15-minute precipitation estimates that can support forecasting and hazard monitoring.

SatWeave combines imagery from six geostationary satellites into a global mosaic with 17 visible, water vapor and infrared channels. It refreshes hourly, and its historical archive spans more than a decade. The dataset underpins WeatherNext 3, enabling the model to produce global predictions more frequently than traditional models that rely on six-hour data refreshes.

PARDIG provides global precipitation estimates at 15-minute intervals and roughly 10-kilometer spatial resolution. Its training used data from NASA’s Global Precipitation Measurement (GPM) Core Observatory, SatWeave imagery and atmospheric fields from ECMWF.

Evaluations compared PARDIG with NASA’s IMERG-Final dataset, using radar networks in the United States, Europe and Japan as well as more than 36,000 weather stations. PARDIG showed more than 30% better detection skill for light-to-moderate precipitation.

SatWeave can support research on solar power forecasting and wildfire monitoring. PARDIG’s potential uses include flood, landslide and drought prediction. Both datasets are available through Google Cloud Storage; PARDIG’s historical catalogs are also available on Google Earth Engine.