DeepMind Opens Up Two Weather Datasets, Including a Rain Map That Beats NASA's

Google DeepMind dropped two foundational datasets on Nov 13, opening up the data plumbing behind WeatherNext 3. PARDIG and SatWeave cover weather, hydrology and climate research, and both are now free for researchers and developers worldwide. The goal, DeepMind says, is to attack the hard problem of watching Earth's weather and water cycles closely enough to predict them.

SatWeave is the eye in the sky. It weaves six geostationary satellites into a single global mosaic, pulling from GOES-East, GOES-West, Meteosat 0, Meteosat Indian Ocean and Himawari. The result: 17 visible, water vapor and infrared channels refreshed hourly, a jump on older composites that carry fewer channels and update less often. WeatherNext 3 leans on that hourly flow, letting it predict globally more often than models stuck on six-hourly refreshes. More than a decade of archives plus live feeds are available, useful for things like solar power forecasting and wildfire monitoring.

PARDIG handles the rain. Full name: Precipitation AI Reanalysis: Densely Inferred GPM. It spits out global precipitation maps every 15 minutes at roughly 10-kilometer resolution, built on AI trained with NASA's Global Precipitation Measurement Core Observatory data, SatWeave imagery and ECMWF atmospheric fields. That matters most where ground radar is thin. Tested against radar networks in the US, Europe and Japan and more than 36,000 weather stations, PARDIG beat NASA's IMERG-Final dataset, lifting detection skill for light-to-moderate precipitation by over 30%.

Both datasets live on Google Cloud Storage, and PARDIG's historical catalogs also sit on Google Earth Engine. DeepMind frames the release as a handoff to the global research community, aimed at sharper models for floods, landslides and droughts. It's the latest piece of Google's wider Earth AI portfolio, which turns AI systems loose on environmental problems.