PARDIG and SatWeave Open as Experimental Weather Data

Google DeepMind on 13 November 2026 said it is releasing two datasets that sit behind its WeatherNext 3 model: SatWeave, a global mosaic of geostationary satellite imagery, and PARDIG, an AI-based global precipitation dataset. The announcement is in a DeepMind blog post by the WeatherNext team, with a companion Google for Developers guide to PARDIG and a paper on arXiv (2611.13941). Historical archives and live feeds of both datasets are being made available. WeatherNext 3 itself was introduced in a DeepMind post dated September 2026, which this announcement treats as already released; the new decision is to open the datasets that helped train and feed it, not to launch the model again.

The names need unpacking. SatWeave is “a global mosaic of geostationary satellite imagery” that, the blog says, “updates hourly” and comes with “a historical archive spanning over a decade.” PARDIG stands for Precipitation AI Reanalysis: Densely Inferred GPM. It is “trained directly on sparse spaceborne radar observations from the NASA Global Precipitation Measurement (GPM) Core Observatory,” and produces “a highly accurate global precipitation map at 15-minute temporal resolution and roughly 10-kilometer resolution, updated with an average latency of about five hours.” The blog also says DeepMind is releasing “historical archives of PARDIG data from 1999 to 2026.”

What PARDIG and SatWeave contain, and how they may be used

The blog names five geostationary satellites whose imagery SatWeave combines: GOES-East and GOES-West, Meteosat 0, Meteosat Indian Ocean, and Himawari. Different projections, formats, sensors and scan schedules, it says, make a coherent global map hard to assemble; SatWeave is DeepMind’s attempt at a unified, hourly view, which WeatherNext 3 uses to produce new global predictions more often than models initialised on analysis data refreshed every six hours. The company says it has used SatWeave in medium-range forecasts and wants others to explore applications such as solar power and wildfire monitoring.

PARDIG, built on an architecture the blog describes as similar to WeatherNext 3, takes SatWeave imagery and atmospheric fields from the European Centre for Medium-Range Weather Forecasts as inputs, alongside the NASA radar training target. The developers’ guide is more precise still. PARDIG is released in two variants under a Creative Commons CC BY 4.0 licence, with no access-request form: PARDIG-NRT, a near-real-time stream from 2016 to the present, and PARDIG-LR, a long-record archive from 1999 to 2023. Live PARDIG-NRT frames on Google Cloud Storage are published in batches every six hours; the guide says effective latency ranges from about two hours to about eight hours and averages about five. Both datasets are on Google Cloud Storage. PARDIG’s historical catalogues are also in Google Earth Engine, with the NRT Earth Engine archive covering January 2016 through October 2026 at launch; the guide says continuous 6-hourly updates should be read from the Cloud Storage Zarr store. The blog does not say that SatWeave’s live feed is in Earth Engine. The two PARDIG variants also use different inputs: NRT takes 0.1° ECMWF HRES fields and SatWeave, while the long record uses coarser 0.25° ERA5 fields, historical GridSat-B1 imagery and polar-orbiting microwave sounders. The guide warns that those input changes, and the evolving satellite mix inside the long record, can introduce non-climatic shifts, so multi-decadal trends should be read with caution.

DeepMind’s own evaluation figures should be read as the company’s. The blog says PARDIG was checked against continental radar networks in the United States, Europe and Japan, and against “over 36,000 weather stations worldwide,” and that it “outperforms leading global baselines across all analyzed rain rates, including NASA’s IMERG-Final.” It claims an improvement of “more than 30%” in hourly precipitation detection skill over IMERG-Final on light-to-moderate precipitation. The developers’ guide adds that, used as a WeatherNext 3 training target, PARDIG achieved “~15% lower probabilistic precipitation forecast error (CRPS) than the IMERG-trained head.” Those are Google’s reported results, not an outside replication.

The guide also notes that PARDIG trains on precipitation retrievals from the Combined Radar-Radiometer Algorithm, or CORRA, on the NASA/JAXA GPM Core Observatory. The blog thanks the public satellite programmes that make the work possible, including ECMWF, NASA, JAXA, NOAA, EUMETSAT, KMA, JMA, KNMI and Synoptic Data.

The good-use case DeepMind itself emphasises is regions that lack dense ground radar. The blog says the company wants partners to explore “improved models for floods, landslides, and drought, especially in regions without dense ground radar networks.”

An experimental dataset is not an official forecast

The developers’ guide is blunt about what PARDIG is not. It calls PARDIG “an experimental AI dataset under active research.” Estimates, it says, are provided for informational and research purposes and are not official weather observations, forecasts, watches or warnings. Users are told to defer to official alerts from their national meteorological service and local emergency authorities, and not to rely on PARDIG as a sole source of information for the protection of life and property.

That caveat is the one that should travel with any reuse of the data, including in the United Kingdom, where DeepMind is based and where the official warnings still come from the Met Office, not from a Google Cloud bucket. Opening the training data behind WeatherNext 3 is a genuine research gift: a decade-scale satellite mosaic, a long precipitation record, and a licence that does not require a special access form. It is not a replacement for a national forecast, and it is not a claim that flood or drought models built on top of it have been operationally approved.

The same caution applies when this release is set beside other AI weather work TSN has covered, from MIT’s climate emulators and the University of Chicago’s AI Weather Almanac to MIT and Microsoft’s subseasonal forecasts and Google’s experimental Earth AI geospatial agent. Those projects are useful context for a growing research stack. None of them, and not this dataset release either, should be treated as an official watch or warning.

The next test is practical and unglamorous: whether independent groups can use the live feeds and the 1999–2026 archives to build, and then validate, flood, landslide or drought models in places where ground radar is thin — and whether they keep Google’s experimental-dataset warning attached when they do.