Google DeepMind has opened two datasets behind its WeatherNext 3 forecasting system, including satellite imagery and AI-estimated rainfall. The useful catch for anyone considering them for live applications: the rainfall estimates arrive, on average, about five hours after the time they describe.
Google DeepMind Watch analysis
What happened
DeepMind released historical records and live feeds for SatWeave, which standardises observations from several weather satellites, and PARDIG, which estimates near-surface rainfall. Both provide data at 15-minute intervals, but PARDIG’s live feed is updated only every six hours, according to XenoSpectrum’s account of the release.
PARDIG-NRT was evaluated against ground radar and rain gauges. The research team says it improved on the IMERG Final precipitation product for light to moderate rain. In one comparison over the contiguous United States, its Critical Success Index was 0.48 against 0.38 for IMERG Final at rainfall of at least 1mm an hour. The team also reports no performance drop in tests on 2024–25 data, outside its training period. These are the researchers’ results; the paper is a preprint, not an independently replicated finding.
Why it matters
A dataset measured in 15-minute slices is not necessarily a feed you can use 15 minutes later. With PARDIG’s roughly five-hour average delivery delay, researchers can explore and test the data, but that timing may rule it out for some operational decisions. SatWeave’s live data has a shorter reported delay of about an hour.
The figures give potential users more to work with than a promising label: there is an evaluation result, a comparison and a delivery schedule. They also make clear why data resolution and data availability are not the same thing.
Our read
This looks like a useful opening of the weather-model toolbox, particularly for researchers who would otherwise have to assemble satellite inputs themselves. But don’t mistake “15-minute rainfall estimates” for a real-time warning service. For decisions involving life or property, official meteorological guidance remains the one to follow.
What to watch
- Whether independent researchers reproduce the reported accuracy results.
- Whether PARDIG’s live-feed update frequency or delivery delay changes.
- How well the estimates perform in regions and conditions where accuracy is limited, including snowfall and mountainous areas.
Discussion spark: Would a rainfall dataset with 15-minute resolution but an average five-hour delivery delay be useful for your work, or does that lag rule out the applications that matter most?
Sources and evidence
- Google DeepMind Releases Satellite and Rainfall Datasets Behind Its WeatherNext 3 Weather AI – XenoSpectrum (14 November 2026, 05:09 UTC)
not affiliated with, endorsed by, or operated by Google or Google DeepMind