DeepMind Details Open Weather Datasets—and the Delay Behind “Near Real Time”

Chinese

A November 13 DeepMind announcement describes public access to PARDIG precipitation estimates and SatWeave satellite mosaics. The technical documentation clarifies their licenses, update cycles and historical coverage, distinguishing dense observations from instant alerts.

AI-generated concept painting: rain clouds, a river delta and a weather station evoke Earth observation and water-cycle research.

A rain map can contain a value for every 15-minute interval and still arrive hours after the rain fell. For researchers building flood or water-management tools, that difference determines which questions a dataset can answer. Coverage, sampling frequency and delivery delay are separate properties.

Google DeepMind's WeatherNext team announced public availability of PARDIG and SatWeave on November 13. They provide precipitation estimates and harmonized satellite imagery used in WeatherNext 3. The news concerns access to the underlying datasets, not a new WeatherNext 3 model launch. Some developer pages carry November 11 update metadata; the announcement alone does not establish the exact first moment the data became publicly accessible. [D1][D2][D4]

PARDIG provides global precipitation estimates on an approximately 10-kilometer grid at 15-minute intervals. Its near-real-time product, PARDIG-NRT, is published on Cloud Storage in batches every six hours. The guide gives an effective delay of roughly two to eight hours, averaging about five hours. A 15-minute timestamp therefore is not a promise of a fresh delivery every 15 minutes. [D3]

That does not make the data unhelpful. It changes where a developer should put it in a workflow. A delayed global estimate might support retrospective analysis, model training or a broader situational picture; it cannot be silently treated as a measurement of conditions at this instant. An operational system should make the age of each observation visible.

The release notes distinguish two precipitation archives: PARDIG-NRT runs from 2016 to the present, while the static long-record product, PARDIG-LR, spans 1999–2023. The guide notes differences in their inputs and cautions about observing-system transitions when combining records. One headline range beginning in 1999 should not be mistaken for a single unchanged live product covering every year. [D2][D3]

SatWeave combines geostationary observations into a common global mosaic with 17 spectral bands. The current stream has hourly timesteps, while the 2016–2025 archive has 30-minute timesteps. The developer documentation lists Cloud Storage access under CC BY 4.0, with a Google Cloud account and no access-request form. Those access terms do not remove a project's separate storage, processing or operational costs. [D2][D4]

DeepMind reports evaluating PARDIG against regional radar networks and more than 36,000 weather stations. These remain developer-reported comparisons; this article does not independently reproduce them. The SatWeave guide describes an experimental research dataset and explicitly says it does not constitute official observations, forecasts, watches or warnings. [D1][D4]

The practical value of the opening is that more teams can inspect and test a common input. A credible downstream study would document the product version, observation age and local validation, then compare its contribution against an existing method. Broader access creates the opportunity for that evidence; it does not establish better flood decisions in every region.

For the global research community, the release is most useful as infrastructure with stated limits. The next consequential results will be applications that explain not only where these data improve a prediction, but also where delay, spatial scale or sparse independent measurements limit what can responsibly be inferred.

Sources:

  1. [D1] Google DeepMind public announcement | November 13, 2026

  2. [D2] WeatherNext release notes | November 2026

  3. [D3] PARDIG developer guide | checked November 14, 2026

  4. [D4] SatWeave developer guide | updated November 11, 2026; checked November 14