This week: Google DeepMind opened up the satellite and rainfall data that feed its WeatherNext 3 forecasts, free for researchers and developers to use.

Elsewhere: How an AI learned to see rain from a single satellite’s narrow stripe, what the data still gets wrong, and three fresh AI drops worth a skim.

The New Prime Signal

Weather apps are great at telling you it rained. Telling you exactly how much fell, everywhere on the planet, a few hours after it happened is a much harder trick, especially in places with no radar and few rain gauges. On Friday, Google DeepMind released two datasets aimed at that gap. They're the same ingredients behind its WeatherNext 3 forecasting model, and the company is handing out both the historical archives and the live feeds. Usually the recipe stays in the kitchen. This time it's taped to the fridge.

What shipped: The first dataset, SatWeave, stitches together imagery from five geostationary satellite positions (GOES-East, GOES-West, two Meteosat positions, and Himawari) into one global picture across 17 visible, water vapor, and infrared bands. Those satellites use different map projections, file formats, and scan schedules, which is why DeepMind says most existing global composites offer four or fewer channels or update only every few hours. SatWeave updates hourly and comes with a historical archive going back to 2016. The second, PARDIG (short for Precipitation AI Reanalysis: Densely Inferred GPM), is an AI-made global rain map at 15-minute steps and roughly 10-kilometer resolution, with archives back to 1999. Both are published under a CC BY 4.0 license, and Google's docs say anyone with a Google Cloud account can read them without filling out an access request form.

Why it matters: Rain decides a lot: crops, water supplies, floods, droughts. Ground radar and dense gauge networks measure it well, but they don't cover every region. PARDIG learns from satellites instead, which is the whole point. DeepMind tested it against ground radar networks in the U.S., Europe, and Japan plus more than 36,000 weather stations, and says it beats leading global baselines at every rain rate it analyzed, including NASA's research-grade IMERG-Final. The headline number: more than 30% better hourly detection of light to moderate rain than IMERG-Final, and DeepMind says it's especially good at catching extreme downpours. Its blog shows PARDIG tracking the July 2025 Texas Hill Country flash floods next to ground radar.

The fine print: Google labels PARDIG an experimental dataset, and its developer guide is upfront about the weak spots. Estimates beyond about 67 degrees north or south lean on extrapolation, because the radar satellite it learned from doesn't fly over those latitudes. Its edge over gauge-adjusted products shrinks over big mountain ranges like the Alps and the Rockies, it tends to underestimate wintertime snow, and it caps rain rates at 75 millimeters an hour, which can clip the most intense storms. The live feed also isn't instant: new data lands in batches every six hours, about five hours behind on average. And Google is clear that it's not an official forecast or warning. For anything involving life and property, listen to your national weather service. The New Prime Signal take: Plenty of AI weather announcements boil down to ‘our model is better, trust us.’ Releasing the training data, the live feed, and the research paper lets other people check that claim and build things DeepMind never thought of. That's a more useful kind of bragging.

The cleverest part is the teacher. NASA and Japan's JAXA run the GPM Core Observatory, a satellite carrying a precipitation radar that measures rain very well but only along a narrow path as it orbits. That's a lot of excellent data with a lot of gaps in between. PARDIG's job is to fill in the gaps. It's trained on those sparse radar readings and learns to estimate rain everywhere else from what the other instruments can see.

How it works: According to DeepMind's post and the developer docs, the pieces fit together like this. The eyes: SatWeave supplies a near-live, hourly view of clouds and storms from the geostationary satellites, which sit over the same spot and watch weather develop. The context: Atmospheric data from ECMWF, the European weather center, adds the state of the air around those clouds. The answer key: The GPM radar readings act as the ground truth PARDIG learns from. The model uses an AI architecture similar to WeatherNext 3's. That loop feeds back into forecasting. WeatherNext 3, which Google launched in September, makes a new forecast every hour using SatWeave, and PARDIG is one of the rainfall targets it trains on. Google's docs say the PARDIG-trained rain forecasts had about 15% lower error than a version trained on NASA's IMERG at early lead times, measured against independent ground radar. Better rain maps make better forecasts. Weather is a team sport.

Try it: The PARDIG guide and SatWeave guide include folder layouts, Python examples, and starter notebooks. Both datasets live on Google Cloud Storage, and PARDIG's historical archives also show up in the Earth Engine data catalog if you'd rather explore on a map. DeepMind says it wants help applying the data to floods, landslides, droughts, solar power, and wildfire monitoring. Fair warning: the 2016–2025 PARDIG archive alone is about 9 TB, so maybe don't download the whole planet onto your laptop.