Google DeepMind opens precipitation and satellite datasets behind its AI weather models
Google DeepMind says it is opening up the foundational precipitation and geostationary satellite datasets behind its AI weather models.
Google DeepMind is releasing two datasets, SatWeave and PARDIG, which it says feed the precipitation and cloud inputs of its recently released WeatherNext 3 weather model. Historical archives and live feeds of both are available to outside researchers. The live and archived SatWeave and PARDIG data are openly available on Google Cloud Storage, and PARDIG's historical catalogs are also in Google Earth Engine. The arXiv preprint describing PARDIG adds licence terms the blog post leaves out: both PARDIG datasets are openly available under a CC BY 4.0 license.
SatWeave deals with a plumbing problem before any modelling begins. It combines imagery from five satellite programs: GOES-East and GOES-West, Meteosat 0 and Meteosat Indian Ocean, and Himawari. Each uses different map projections, file formats, sensors and scan schedules. Scans run every 10 to 30 minutes, which makes a single coherent global map hard to build. DeepMind says most existing global composites either drop data so they offer four or fewer channels, or they update only every few hours. SatWeave covers 17 visible, water vapor and infrared channels and updates hourly, while models initialized on analysis data refresh every six hours. The post's caption says SatWeave blends observations from six geostationary satellites, but the body names five, and the post doesn't reconcile the two. The historical archive spans over a decade.
PARDIG, which stands for Precipitation AI Reanalysis: Densely Inferred GPM, rests on a sparser source. NASA's Global Precipitation Measurement Core Observatory carries a spaceborne radar, but its observations are sparse. DeepMind trains a model on those radar observations, using SatWeave imagery and atmospheric fields from ECMWF as inputs. The aim is for the model to learn to predict rain where the radar was not looking, and then apply it densely across the globe. The output is a precipitation map at 15-minute temporal resolution and roughly 10-kilometer resolution, with an average latency of about five hours. The paper gives the spatial grid as 0.1 degrees.
DeepMind says PARDIG was evaluated against three continental radar networks in the US, Europe and Japan, and against more than 36,000 weather stations worldwide. The headline figure is that PARDIG improves hourly precipitation detection skill over IMERG-Final, a NASA research-quality dataset, by more than 30 percent on light-to-moderate precipitation. That is measured with the Critical Success Index for 2016 to 2023. The post says PARDIG is particularly effective at capturing extreme precipitation. Its example is hourly rainfall during the July 2025 Texas Hill Country flash floods, comparing NOAA's ground radar at about 1 km with PARDIG and IMERG-Final, both at about 10 km. The example sets a 1 km radar beside two 10 km products, and the post gives no score for that event.
The company's broadest claim is that, because PARDIG learns from satellites rather than relying solely on ground measurements, it should be accurate in places without ground data: 'we expect it to have reliable accuracy everywhere on Earth.' That is an expectation, not a result. The evaluations in the post use ground radar and weather stations, which are not always available or well covered, and the post doesn't describe a test in regions without dense ground networks. The arXiv abstract lists a wider set of comparisons, saying PARDIG consistently outperforms IMERG V07 Final and Early, MSWEP V3, ERA5 and HRES, with the largest improvements at extreme intensities. That is a preprint, and the abstract makes these claims without the post's extra material.
The datasets depend on public satellite and reanalysis work. The post's acknowledgements credit teams across Google DeepMind and Google Research, working with Google Cloud and Google Geo. They also thank ECMWF, NASA, JAXA, NOAA, EUMETSAT, KMA, JMA, KNMI and Synoptic Data for the open satellite programs and datasets that make the work possible. Google Earth Engine, where PARDIG's historical catalogs are hosted, is described on its own page as now available for commercial use and free for academic and research use.
What changed
Google DeepMind is releasing the SatWeave and PARDIG datasets, which feed the precipitation and cloud inputs of WeatherNext 3, with historical archives and live feeds available to outside researchers.
Type: New capability · Applies to: SatWeave, PARDIG, WeatherNext 3, Google Cloud Storage, Google Earth Engine
Why it matters
Groups building forecasting or climate-related models may be able to use these datasets as training or evaluation inputs. The announcement does not say how access works, so its practical value is still unclear.
What you can do
Check DeepMind's post for the access route and licence terms before planning any use of the datasets.
Original source
Where we saw it
Google DeepMind
AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works