DeepMind Opens PARDIG and SatWeave: Free Rain and Satellite Data Behind WeatherNext 3
Google DeepMind opened two of the datasets behind its WeatherNext 3 forecasting model to everyone on 13 November 2026 — three days ago — and they are free to use. PARDIG is a global rainfall map updated every 15 minutes at roughly 10-kilometre resolution, with records reaching back to 1999. SatWeave is an hourly global mosaic of satellite imagery across 17 channels. Both are released under a Creative Commons Attribution (CC BY 4.0) licence, with live feeds and historical archives on Google Cloud Storage.
What DeepMind released on 13 November
A foundational dataset is a cleaned, harmonised base layer that other models and applications are built on. Raw satellite data is messy: different satellites use different map projections, file formats, sensors and scan schedules, and stitching them into one coherent global picture is a major engineering job. DeepMind has done that job twice, and is now publishing the results instead of keeping them in-house.
The announcement, published on the Google DeepMind blog on 13 November 2026 by the WeatherNext team, is short and specific. "Today, we are releasing two foundational datasets, SatWeave and PARDIG, to support the global community in building weather, hydrology, and climate applications," it says. "Both datasets are key to the advances in our recently released WeatherNext 3 model." The live feeds and historical archives are openly available on Google Cloud Storage, and PARDIG's historical catalogues are also available in Google Earth Engine. There is no access-request form. Anyone with a Google Cloud account can get the data.
![]() |
| AI-generated illustration (generated 16 Nov 2026). |
SatWeave: one hourly picture of the whole planet
SatWeave is the satellite half of the release. It combines imagery from five named geostationary satellite systems that cover different parts of the globe: GOES-East and GOES-West, operated by the U.S. National Oceanic and Atmospheric Administration; Meteosat 0 and Meteosat Indian Ocean, operated by Europe's EUMETSAT; and Himawari, operated by Japan. Because each uses different projections, formats, sensors and scan schedules — ranging from 10 to 30 minutes — forming a coherent global map from them is difficult, and most existing global composites drop incompatible aspects, offering four or fewer channels or updating only every few hours.
SatWeave blends those observations into a single global mosaic across 17 visible, water-vapour and infrared channels. It updates hourly, and that hourly refresh is what allows WeatherNext 3 to produce new global predictions every hour, in contrast to models initialised on analysis data refreshed every six hours. DeepMind is releasing the hourly live feed plus a historical archive spanning over a decade. The practical uses go beyond weather forecasting: DeepMind points to solar power forecasting, where cloud cover over the next hours determines output, and wildfire monitoring, where a live view of developing smoke and heat matters.
PARDIG: rain, everywhere, every 15 minutes
PARDIG is the rainfall half — and the more striking of the two. Its full name is Precipitation AI Reanalysis: Densely Inferred GPM. It is an AI-based global precipitation dataset, built on an architecture similar to WeatherNext 3, and trained directly on sparse spaceborne radar observations from the NASA Global Precipitation Measurement (GPM) Core Observatory, using SatWeave imagery and atmospheric fields from the European Centre for Medium-Range Weather Forecasts (ECMWF) as inputs.
The result is a global precipitation map at 15-minute temporal resolution and roughly 10-kilometre spatial resolution, updated with an average latency of about five hours. In plain terms: for any point on Earth, PARDIG estimates how hard it rained in each 15-minute slice, and that estimate reaches users roughly five hours after the rain fell. That is far faster than NASA's research-grade IMERG-Final product, which takes months, and it covers the whole planet — including the many regions with no dense ground radar or rain-gauge network at all. Because PARDIG learns directly from satellites rather than relying solely on ground measurements, DeepMind says it expects reliable accuracy everywhere on Earth, even where ground measurements are unavailable.
| AI-generated illustration (generated 16 Nov 2026). |
How good is PARDIG? DeepMind's own numbers
DeepMind evaluated PARDIG against three continental radar networks across the United States, Europe and Japan, plus more than 36,000 weather stations worldwide. Its headline result: PARDIG improves hourly precipitation detection skill over NASA's IMERG-Final by more than 30% on light-to-moderate precipitation, and outperforms leading global baselines across all analysed rain rates. It also proved particularly effective at capturing extreme precipitation, which existing global products often struggle to detect.
Those figures are DeepMind's own, from its own evaluation, and no independent replication has been published yet. That does not make them wrong — the evaluation covers three continents and tens of thousands of stations, which is a serious test — but it does mean they should be reported as DeepMind's claim, not as settled fact. The underlying research was posted as an arXiv preprint on 10 November 2026, three days before the blog post, and independent researchers will now be able to test the data themselves, because it is public. That is the point of opening it.
It is also worth keeping this release in perspective. Open datasets are one of the quieter ways the big AI labs shape the field: whoever supplies the base layer influences what everyone else builds. Our explainer of the State of AI Report 2026 tracked how open models and open data have become strategic assets rather than charity, and PARDIG and SatWeave fit that pattern — genuinely useful, genuinely free, and also good for Google's position at the centre of the weather-AI ecosystem.
PARDIG's two archives: read the dates carefully
The blog says DeepMind is releasing "historical archives of PARDIG data from 1999 to 2026." That is the combined span, and it needs a footnote, because PARDIG actually comes in two variants with different date ranges. Google's own developer documentation for the dataset spells out the split, and it matters for anyone planning to use the data.
| Variant | Coverage | What it is |
|---|---|---|
| PARDIG-NRT (near-real-time) | 2016 to present | Live feed, updated every 6 hours, ~5-hour average latency; uses ECMWF HRES fields + SatWeave |
| PARDIG-LR (long-record) | 1999 to 2023 | Static archive; uses ERA5 reanalysis + NOAA GridSat-B1 + passive-microwave sounders |
| SatWeave | Over a decade + hourly live feed | 17-channel global mosaic from geostationary satellites |
What you can build with it — and what you cannot
The honest uses are the ones DeepMind names: weather, hydrology and climate applications, especially flood, landslide and drought modelling in regions without dense ground radar. A developer in a country with sparse rain gauges can now pull a global rainfall record into a flood-warning prototype in an afternoon. A solar operator can feed SatWeave's hourly cloud mosaic into a generation forecast. A climate researcher can train a precipitation model without first spending months harmonising satellite records.
The release also sits inside a bigger Google story. The same week has seen flagship-model competition intensify — our report on Gemini 4 Carbon covered Google's newest flagship — and open research datasets are part of how Google keeps developers inside its orbit even as the model race tightens. The same logic applies to AI in science more broadly: our coverage of Claude AI drug discovery showed frontier labs opening scientific tools to researchers, and the pattern repeats here. Open data is good for science and good for the lab that opens it.
| AI-generated illustration (generated 16 Nov 2026). |
Why this release matters beyond weather
Whether PARDIG lives up to DeepMind's 30%-plus claim will be settled by independent testing now that the data is public. That is the healthy way round: publish the data, let the community check it. And the funding backdrop matters too — even as markets argue over is the AI bubble bursting, Google is spending on open scientific infrastructure that pays off over years, not quarters. Datasets like these are the slow, unglamorous part of the AI boom — and they may outlast most of the hype.
Key Takeaways
- DeepMind released PARDIG and SatWeave on 13 November 2026 — three days before this article — free under CC BY 4.0 on Google Cloud Storage.
- PARDIG maps global rainfall every 15 minutes at about 10 km resolution, with about 5 hours of latency; SatWeave is an hourly 17-channel satellite mosaic.
- Archive span: PARDIG-NRT covers 2016 to the present, PARDIG-LR covers 1999 to 2023; the blog's "1999 to 2026" is the combined span.
- DeepMind's own tests claim PARDIG beats NASA's IMERG-Final by over 30% on light-to-moderate rain; no independent replication yet.
- The data is experimental and as-is — useful for research, not a substitute for official weather warnings.
Conclusion
PARDIG and SatWeave matter more over a year than a day. Free, open, global rainfall and satellite data — the ingredients behind WeatherNext 3 — now belong to anyone with a cloud account. The 30%-plus claim needs independent checking, and the archive dates need reading carefully. The next flood model built on this data will show whether the gift was as good as it looks.
Sources
- Google DeepMind — PARDIG & SatWeave: Foundational Datasets for Earth Science (13 Nov 2026)
- Google for Developers — PARDIG, WeatherNext guides (licence CC BY 4.0; PARDIG-NRT / PARDIG-LR split)

Comments
Post a Comment