Google opens the satellite and rain data behind WeatherNext 3

Google DeepMind is releasing two datasets that feed its WeatherNext 3 forecasting model: PARDIG, a global AI-built precipitation record, and SatWeave, an hourly mosaic of geostationary satellite imagery. In its announcement on X, DeepMind called them “the foundational datasets behind WeatherNext 3” and said it wants to help researchers.

The details come from DeepMind’s blog post, which says both datasets are openly available on Google Cloud Storage as live feeds and historical archives.

Two datasets, one 15-minute rain record

SatWeave stitches imagery from five geostationary satellites into a global mosaic: GOES-East, GOES-West, Meteosat 0, Meteosat Indian Ocean and Himawari. A video caption on the same page mentions six, so the exact count is unclear. The mosaic spans 17 visible, water vapor and infrared channels. Source scans arrive every 10 to 30 minutes depending on the satellite, and the mosaic updates hourly. DeepMind contrasts that with models that start from analysis data refreshed every six hours. The archive covers more than a decade. According to the post, SatWeave feeds the medium-range forecasts of WeatherNext 3.

PARDIG, short for Precipitation AI Reanalysis: Densely Inferred GPM, is a precipitation dataset at 15-minute and roughly 10 km resolution. It is trained on sparse spaceborne radar observations from NASA’s GPM Core Observatory, with SatWeave imagery and ECMWF atmospheric fields as inputs. Average latency is about five hours. History runs from 1999 to 2026, and a live feed is included. DeepMind says it was used to train the precipitation forecasts of WeatherNext 3.

Both datasets can be downloaded from Google Cloud Storage. PARDIG’s historical catalogs are also in Google Earth Engine, and a paper is on arXiv (abs/2611.13941). WeatherNext developer documentation is at developers.google.com/weathernext.

A claimed 30% gain over NASA’s rain benchmark

DeepMind says it checked PARDIG against three continental radar networks (US, Europe and Japan) and more than 36,000 weather stations worldwide. Its headline claim: “PARDIG improves hourly precipitation detection skill over IMERG-Final by more than 30% on light-to-moderate precipitation.” The post says it also beat baselines, including IMERG-Final, across all analyzed rain rates and did well on extreme precipitation. These are DeepMind’s own evaluations; the post does not say they were independently reproduced, and the peer-reviewed status of the paper is not stated.

The aim is global coverage, including regions without ground radar or rain gauges, where the post says the model should remain accurate.

How PARDIG compares with NASA’s IMERG

IMERG, NASA’s merged precipitation product, is the natural yardstick. According to NASA’s GPM page, IMERG updates every half hour, and its Early Run arrives with about 4 hours of latency. The Final Run, the research-grade version that DeepMind compares against, takes roughly 2.5 to 3.5 months (the page gives both figures). PARDIG’s reported five-hour latency, at 15-minute steps, therefore puts a live, finer-timed record next to a benchmark that arrives months later. That is the comparison of interest, though DeepMind does not say whether its skill figures were computed against the Final Run on a like-for-like timeline.

NASA also notes that IMERG’s continuous record starts in 1998 and that its infrared-based estimates are limited to 60°N-60°S. PARDIG advertises global coverage and a record from 1999, so the two overlap almost entirely in time. Whether PARDIG holds up at high latitudes, where satellite geometry is poor, is not addressed in the post.

What researchers can do with it, and what is missing

For forecasters and hydrologists, a rain record that is both historical and live, at 15-minute steps, can be used to train or test their own models without building a satellite pipeline. SatWeave offers the same convenience for imagery. Google’s own infrastructure is not the point here; the point is that the training data behind a flagship forecast model is now inspectable.

The post does not state the licenses, file sizes or download costs, so reuse terms, including for commercial work, need checking in the developer documentation. It also does not say how WeatherNext 3 performs without these inputs, so the datasets’ contribution to the model’s accuracy rests on DeepMind’s description. The post thanks ECMWF, NASA, JAXA, NOAA, EUMETSAT, KMA, JMA, KNMI and Synoptic Data for their contributions.