Felix Laurent
Google DeepMind has released PARDIG and SatWeave, the two datasets behind its WeatherNext 3 weather model. The lab announced the release on November 13, 2026. Both come with a live feed and a historical archive on Google Cloud Storage. Researchers can build their own weather, flood and climate tools on top.
PARDIG estimates rainfall across the whole planet every 15 minutes. SatWeave stitches geostationary satellites into one hourly view of Earth's clouds and storms. DeepMind says the rainfall data detects precipitation more than 30% better than NASA's leading global product.
Why DeepMind is releasing weather data
Every forecast starts with a picture of the atmosphere right now. Satellites and ground stations measure the planet constantly, but turning those raw readings into one consistent global map quickly is hard. The WeatherNext team wrote that predicting weather "starts with how well we can observe the Earth right now."
That picture already powers Google's own forecasts. WeatherNext 3, released on September 3, makes hourly forecasts at 5-kilometer resolution, against 25 kilometers for WeatherNext 2. It runs in Google Search, the Gemini app and Google Maps. Opening its inputs lets others train and test models on the same data, starting with the satellite view.
What SatWeave does
SatWeave combines imagery from geostationary satellites that each watch a different part of the globe. DeepMind's blog names five: GOES-East and GOES-West, Meteosat 0° and Meteosat Indian Ocean, and Himawari. A figure caption on the same page says six, a detail DeepMind has not clarified.
Merging them is harder than it sounds. Each satellite uses its own map projection, file format, sensors and scan schedule, from every 10 to every 30 minutes. Most existing global composites drop what does not fit, keeping four or fewer channels or updating only every few hours. SatWeave keeps 17 visible, water vapor and infrared channels.
The mosaic updates every hour. That lets WeatherNext 3 start a new global forecast each hour, while many models start from analysis data refreshed every six hours. DeepMind is releasing the hourly live feed and an archive covering more than a decade. It wants others to try it for solar power forecasts and wildfire monitoring. SatWeave also feeds the second dataset.
What PARDIG measures
Rain is one of the hardest things to track globally. Ground radar and dense rain gauge networks give good data, but much of the world, including large parts of Africa, has little of either. PARDIG, short for Precipitation AI Reanalysis: Densely Inferred GPM, tries to fill that gap from space.
The model behind it uses an architecture similar to WeatherNext 3. It was trained on spaceborne radar readings from NASA's Global Precipitation Measurement Core Observatory. As inputs, it uses SatWeave imagery and atmospheric data from the European weather center ECMWF. The result is a global rainfall map every 15 minutes at about 10-kilometer resolution, available about five hours after the fact.
DeepMind tested PARDIG against ground radar networks in the US, Europe and Japan and more than 36,000 weather stations. It says PARDIG beat leading global products at every rain rate, including NASA's IMERG-Final. For light to moderate rain, it improved hourly detection by more than 30%. DeepMind says it also does well on extreme rainfall, which global products often miss.
One test case was the July 2025 flash floods in Texas Hill Country. DeepMind compared hourly rainfall from NOAA's ground radar, PARDIG and IMERG-Final over the same storm.
The historical archive runs from 1999 to 2026, alongside the live feed. Because PARDIG learns from satellites, DeepMind expects it to stay accurate even where no ground measurements exist. That is where the data could matter most.
How researchers can use PARDIG and SatWeave
Both datasets are open on Google Cloud Storage, and PARDIG's historical catalogs are also in Google Earth Engine. DeepMind lists several uses:
- Flood warnings in regions without ground radar
- Landslide and drought models built on long rainfall records
- Wildfire smoke tracking from the hourly satellite view
- Solar power forecasts based on cloud cover
The datasets join Google's Earth AI portfolio. The work came from Google DeepMind and Google Research with Google Cloud and Google Geo. It relies on open satellite programs from ECMWF, NASA, JAXA, NOAA, EUMETSAT and national weather agencies in Korea, Japan and the Netherlands. The reaction showed which of those uses people care about most.
How the community reacted
Many replies under DeepMind's post focused on places without radar. One user wrote that a free rainfall feed every 15 minutes "is honestly a bigger deal than the forecast model itself" for the global south. Another pointed out that Accra in Ghana floods almost every June with little warning.
Researchers also asked practical questions. Several wanted to know whether DeepMind provides a fixed held-out test period, so outside models can be compared with WeatherNext 3 without data leakage. Others asked whether the rainfall gains hold for the heaviest storms and whether uncertainty estimates are included. DeepMind had not answered those questions in the thread.
Why PARDIG and SatWeave matter
For weather researchers, this is a rare release of the data behind a production AI forecast. A global rainfall record every 15 minutes back to 1999 is a strong base for flood and drought research. It matters most in regions that weather services cannot cover with radar.
The datasets still need outside checks. DeepMind ran the main evaluations itself, and the satellite count in its own post does not match. Independent tests on extreme storms and clear benchmark splits will decide how widely forecasters adopt PARDIG and SatWeave. If they hold up, the biggest winners could be flood-prone countries that never had a radar network.
Sources
- Opening the datasets behind our AI weather models, Google DeepMind – November 13, 2026
- PARDIG and SatWeave announcement, Google DeepMind on X – November 13, 2026
- Introducing WeatherNext 3, Google – September 2026
- WeatherNext, Google DeepMind
- Google WeatherNext 3 delivers hourly forecasts at 5-kilometer resolution, TechRepublic – September 2026
