INDEPENDENT PUBLICATION · AMERICAN AMBITION BECOME A FOUNDING SPONSOR NEWS & ANALYSIS / DATED AT PUBLICATION

SCIENCE / LATEST COVERAGE

A clearer view.
More room to discover.

Google is opening two datasets behind its weather models, giving researchers more building blocks for Super Intelligence (SI) tools that study rain, storms and the water cycle.

Open data, practical possibilities.

A better weather tool starts with a better picture of what is happening now. That picture is surprisingly difficult to assemble: satellites see different parts of Earth, instruments measure different things, and the observations do not all arrive on the same schedule.

On November 13, Google DeepMind released SatWeave and PARDIG, two datasets used in its WeatherNext 3 forecasting model. The release opens historical archives and live feeds so researchers and developers can build their own weather, water and climate applications.

The useful advance is access. Instead of every research team having to start by stitching together raw observations, the new release gives them shared ingredients to work with. That can leave more room for the science and engineering that turn data into a useful tool.

A shared view from space.

SatWeave combines geostationary satellite imagery into a global mosaic. These satellites watch regions of Earth from fixed positions in the sky, but their instruments, formats and observation schedules differ.

Google says SatWeave brings together 17 visible, water-vapor and infrared channels, with an hourly live feed and a historical archive spanning more than a decade. The aim is a more coherent view of clouds and developing storms.

The company identifies solar-power forecasting and wildfire monitoring as areas it hopes to explore with the research community. Those are potential applications of the data, rather than newly completed services announced in this release.

Putting rainfall on the map.

PARDIG focuses on precipitation. It learns from spaceborne radar measurements from NASA’s Global Precipitation Measurement (GPM) observatory, using satellite imagery and atmospheric information to estimate rainfall across the globe. The dataset’s full name is Precipitation AI Reanalysis: Densely Inferred GPM.

Google describes maps with roughly 10-kilometer spatial resolution and 15-minute time steps. There is an important distinction between that time spacing and delivery speed: the company reports an average delay of about five hours before updates are available.

The team’s research preprint, submitted November 10, describes evaluation against radar networks in the United States, Europe and Japan, and more than 36,000 surface stations worldwide. The authors report improved precipitation estimates against the comparison datasets they studied. That is a research result reported by the team, not an ASIndex assessment.

The release includes historical PARDIG data reaching back to 1999. A longer record lets researchers study past events as well as develop tools using recent observations.

What researchers can build.

Google says both datasets’ live and historical feeds are openly available on Google Cloud Storage, with PARDIG’s historical catalogs also available in Google Earth Engine. Its WeatherNext developer documentation is a starting point for exploring the wider forecasting tools.

The researchers want to enable work on floods, landslides and drought, particularly where ground-based measurement networks are sparse. Shared data can help teams investigate those questions without claiming that a new warning system is already operating everywhere.

Our reading: this is a constructive example of American SI research widening the workshop door. The headline is not just a new model. It is more people getting useful materials to build with—and more chances to turn a clearer picture of the planet into something practical.

Keep exploring

Useful science. Shared possibilities.

Explore American innovation ↗