Google DeepMind opens up its key weather datasets SatWeave and PARDIG

Google DeepMind has opened up the two core datasets behind its WeatherNext 3 weather model, SatWeave and PARDIG, giving researchers worldwide real-time data streams and historical archives for weather, hydrology, and climate modeling.

SatWeave combines data from five geostationary weather satellites, including GOES, Meteosat, and Himawari. It updates every hour and covers more than ten years of history, letting WeatherNext 3 produce hourly global forecasts from live cloud and storm images, while traditional data usually refreshes every six hours.

PARDIG is an AI-based global precipitation dataset with a spatial resolution of about 10 km and 15-minute updates. Tests against radar networks in the US, Europe, and Japan, plus 36,000 weather stations, show it beats mainstream products like NASA IMERG-Final, with over 30% better detection of light to moderate rain, and it stays reliable in areas without ground radars.

The archives and live streams are hosted on Google Cloud Storage under the CC BY 4.0 license, and can be used for things like solar power forecasting and wildfire monitoring.