Knowing how much rain is falling right now, everywhere, is harder than it sounds. Rich countries have dense radar and rain-gauge networks; much of the world does not, and that is often where floods do the most damage. Google DeepMind has released two datasets that sit behind its WeatherNext 3 forecasting model, a global satellite mosaic called SatWeave and an AI-built rainfall record called PARDIG, with historical archives and live feeds that anyone can use.
SatWeave solves a plumbing problem. Geostationary weather satellites, including GOES-East and GOES-West, two Meteosat satellites and Himawari, each use different map projections, file formats, sensors and scan schedules, ranging from 10 to 30 minutes. Most global composites cope by dropping detail, offering four or fewer channels or updating every few hours. SatWeave blends them into a single global mosaic with 17 visible, water-vapour and infrared channels, updated hourly, with a historical archive of more than a decade. That hourly refresh lets WeatherNext 3 start new forecasts more often than models that rely on analysis data refreshed every six hours.
PARDIG is the more ambitious piece. It is trained on sparse readings from the spaceborne radar on NASA's Global Precipitation Measurement Core Observatory, using SatWeave images and atmospheric fields from ECMWF as inputs, and fills in a global precipitation map at 15-minute and roughly 10-kilometre resolution, with an average delay of about five hours. Google evaluated it against radar networks in the US, Europe and Japan and more than 36,000 weather stations, and reports that it beats NASA's research-grade IMERG-Final across all rain rates analysed, improving hourly detection of light-to-moderate rain by more than 30%, and doing particularly well on extremes. The archive runs from 1999 to 2026.
The practical value is in the openness. Both datasets are on Google Cloud Storage, and PARDIG's history is also in Google Earth Engine. Hydrologists, insurers and disaster agencies working in places without radar could use them for flood, landslide and drought models, and Google suggests SatWeave for solar power and wildfire monitoring. This continues a pattern of AI weather work, from DeepMind and others, that is now as much about data as about models.
The caveats matter. The comparison with IMERG-Final is Google's own, and independent hydrologists will need to confirm it. Ten-kilometre resolution is coarse next to ground radar at about one kilometre, so it will not replace local systems. Google's claim of reliable accuracy everywhere is an expectation based on learning from satellites, and the evaluation regions are exactly the places that already have good radar.