What was released
Both datasets sit underneath WeatherNext 3, DeepMind's current weather model. SatWeave supplies the live picture: a single global mosaic stitched from geostationary satellites that scan on different schedules, between 10 and 30 minutes apart, in different projections and formats. DeepMind says most existing global composites settle for four or fewer channels, or refresh every few hours. SatWeave carries 17 channels and updates hourly, which DeepMind says allows WeatherNext 3 to issue new global predictions off fresh imagery instead of analysis data refreshed every six hours. The historical archive spans over a decade.
PARDIG is the record the model trained its rainfall forecasts on. Ground radar and rain gauges are good where they exist and absent across much of the planet, so PARDIG learns from the sparse radar aboard NASA's GPM Core Observatory, with SatWeave imagery and ECMWF atmospheric fields as inputs. The output is a global precipitation map every 15 minutes at roughly 10 kilometers, arriving with an average latency of about five hours. The archive runs from 1999 to 2026.
The numbers are the author's own
DeepMind evaluated PARDIG against radar networks in the United States, Europe and Japan, and against more than 36,000 weather stations. It reports that PARDIG outperforms leading global baselines, including NASA's IMERG-Final, across every rain rate it analyzed, with the headline figure a detection-skill improvement of more than 30% on light-to-moderate precipitation. The chart behind that claim covers 2016 to 2023.
The announcement presents DeepMind's own evaluation; this article cites no independent reproduction of it. The 30% figure is a vendor measurement of a vendor dataset, published by the team that built both. Anyone citing it before a third party reproduces the comparison is citing developer-reported results.
Count the satellites
The announcement cannot settle on how many satellites SatWeave uses. The introduction says it combines imagery from five satellites and names them: GOES-East, GOES-West, Meteosat at 0 degrees, Meteosat over the Indian Ocean, and Himawari. The SatWeave section, a short scroll later, says it blends observations from six geostationary satellites. No sixth satellite is named anywhere in the post, and the post does not explain whether the two counts refer to different periods or configurations.
It is the kind of detail a release post should get right, because coverage judgments start from which spacecraft are in the mosaic. If the count is six, the archive documentation should say which spacecraft joined and when its record starts.
Two other terms are missing rather than wrong. The post calls both datasets "openly available" and gives their locations, but states no license for either. It also gives no spatial resolution for SatWeave, the figure a developer needs before deciding whether the mosaic can stand in for a regional feed.
Who gets what
The practical value is the archive, not the announcement. A single precipitation record spanning 1999 to 2026, covering regions that have never had dense radar coverage, is a training set flood, landslide and drought modelers can now pull from one place. The hourly SatWeave feed is the narrower offer: most useful to teams already building on Google Cloud, less distinctive where national weather services publish their own composites.
ALFW has not downloaded or run either dataset. All performance figures are DeepMind's own evaluations as published in its announcement.
Sources
- [1] Google DeepMind, "Opening the datasets behind our AI weather models" (November 13, 2026)Read source