DeepMind opens the satellite and rainfall data behind WeatherNext 3, a gift for places without radar

Google DeepMind opened two datasets that feed its WeatherNext 3 forecasting model (our WeatherNext 3 coverage). SatWeave stitches geostationary satellite imagery into an hourly global mosaic across 17 channels, merging GOES-East, GOES-West, Meteosat and Himawari satellites that use different projections, formats and scan schedules of 10 to 30 minutes. DeepMind says most existing global composites offer four or fewer channels or update only every few hours. PARDIG is an AI-built global precipitation record at roughly 10-kilometer, 15-minute resolution. Both come with live feeds and historical archives, PARDIG going back to 1999.

Hourly global satellite views let WeatherNext 3 update more often than models tied to analysis data refreshed every six hours.

Rain is hard to measure across most of the planet. Ground radar and dense rain-gauge networks are concentrated in wealthier regions. A free, consistent global rainfall picture could help researchers and governments working on floods, landslides and drought where data has always been thinnest. Models get the headlines, but data lets other people build. With live and historical data on Google Cloud Storage and PARDIG archives in Earth Engine, DeepMind is inviting hydrologists, energy forecasters and wildfire researchers to do work it won’t do itself. It names solar power and wildfire monitoring as uses for SatWeave. NVIDIA is pursuing a similar open approach with its Earth-2 weather models in Europe (coverage).

DeepMind says PARDIG beats leading global baselines, including NASA’s IMERG-Final, at every rain rate tested, with more than 30 percent better hourly detection of light-to-moderate rain and notable gains on extreme events. It evaluated against radar networks in the U.S., Europe and Japan plus more than 36,000 weather stations. Those results deserve to be taken seriously, and independent evaluations are still needed, especially where the dataset should help most and ground truth is scarcest.

PARDIG is trained on sparse spaceborne radar from NASA’s GPM mission, so DeepMind expects it to hold up everywhere. Expecting isn’t proving. If emergency planners in data-poor countries lean on an AI-inferred rainfall product, they need to know where it is weak. A model that misses a flash flood where no gauge can catch the error is a real risk.

This work stands on public satellite programs and agencies, including NASA, NOAA, EUMETSAT, JMA and ECMWF. Opening the derived datasets is a fair way to give back. Access needs to stay open and stable, because communities that build warning systems on a feed need it to exist in ten years. Shared data as the real bottleneck is a theme across AI for science, from biology’s virtual cell effort (coverage) to dementia research (coverage). On balance this is a clear plus. More in our research coverage and Google coverage.