Google DeepMind has publicly released two datasets that underpin its WeatherNext 3 forecasting model, it said in a November 13 blog post. SatWeave is a global mosaic of geostationary satellite imagery, and PARDIG is an AI-generated global precipitation record. Live and historical data for both are on Google Cloud Storage, and PARDIG's historical catalog is also on Google Earth Engine.
SatWeave combines imagery from GOES-East, GOES-West, Meteosat 0, Meteosat Indian Ocean and Himawari into a single global product with 17 visible, water-vapor and infrared channels. It updates hourly, compared with the six-hour analysis cycle other models use, and its archive covers more than a decade. WeatherNext 3 uses it for medium-range forecasts.
PARDIG, short for Precipitation AI Reanalysis: Densely Inferred GPM, estimates rainfall worldwide at 15-minute and roughly 10 km (0.1°) resolution, with an average latency of about five hours. The model behind it is a spherical graph-mesh transformer that takes SatWeave imagery and ECMWF atmospheric fields as input. It is trained to predict the precise but sparse measurements from the precipitation radar on NASA's GPM Core Observatory. The near-real-time product starts in 2016, and a historical version goes back to 1999.
DeepMind evaluated PARDIG against radar networks in the US, Europe and Japan over 2016–2023 and against more than 36,000 weather stations. The blog post says it improves hourly detection skill over NASA's IMERG-Final by more than 30% for light-to-moderate precipitation. The accompanying arXiv paper, by Guy Shalev, Stephan Rasp, Boris Babenko and 14 co-authors, reports that PARDIG consistently outperforms IMERG V07, MSWEP V3, ERA5 and HRES, with the largest gains at extreme intensities. It also reports that WeatherNext 3 makes more skillful precipitation forecasts when trained on PARDIG. The paper is a preprint and has not been peer reviewed.
According to the paper, both datasets are available under a CC BY 4.0 license. Data providers acknowledged include ECMWF, NASA, JAXA, NOAA, EUMETSAT, KMA, JMA, KNMI and Synoptic Data. For people building weather, flood, agriculture or climate models, this provides a dense, consistent global rainfall record and hourly satellite input that can be used commercially with attribution. Until now that kind of data took major effort to assemble from separate agency sources.