Skip to content
AI.info

The Pulse

Google DeepMind opens two datasets used by WeatherNext 3

Google DeepMind has released historical archives and live feeds of its SatWeave satellite imagery and PARDIG precipitation datasets. The datasets were used in its WeatherNext 3 forecasting model and are available on Google Cloud Storage.

Google DeepMind opens two datasets used by WeatherNext 3

Researchers and developers can now access live feeds and historical archives of two datasets used in Google's WeatherNext 3 weather forecasting model. Google DeepMind released SatWeave, a mosaic of satellite imagery, and PARDIG, an AI-based precipitation dataset, on November 13.

SatWeave gives the model a view of developing clouds and storms, while PARDIG was used to train its precipitation forecasts. The datasets are open for use in weather, hydrology and climate applications.

Combining satellite views

SatWeave combines imagery from five satellites covering different parts of the globe: GOES-East, GOES-West, Meteosat 0, Meteosat Indian Ocean and Himawari. Their sensors, formats, map projections and scan schedules differ, with scans ranging from 10 to 30 minutes.

The resulting mosaic covers 17 visible, water vapor and infrared channels. DeepMind says SatWeave updates hourly, allowing WeatherNext 3 to produce new global predictions, whereas some models start from analysis data refreshed every six hours.

The archive spans more than a decade, alongside the hourly live feed. DeepMind has used SatWeave in medium-range forecasts and wants to explore other applications with researchers and developers, including solar power and wildfire monitoring.

Mapping precipitation

PARDIG, short for Precipitation AI Reanalysis: Densely Inferred GPM, uses an AI architecture similar to WeatherNext 3. It was trained directly on sparse spaceborne radar observations from NASA's Global Precipitation Measurement Core Observatory, using SatWeave imagery and atmospheric fields from ECMWF as inputs.

The precipitation maps have 15-minute temporal resolution and roughly 10-kilometer spatial resolution, with an average update latency of about five hours. The historical archive covers 1999 to 2026, and a live feed is also being released.

DeepMind developed the dataset in part to address gaps in ground radar and rain gauge coverage. It expects PARDIG to have reliable accuracy worldwide, including in areas without ground-based measurements, because the system learns directly from satellite observations.

Testing and access

The team evaluated PARDIG against radar networks in the United States, Europe and Japan, as well as more than 36,000 weather stations worldwide. Google DeepMind reports that PARDIG outperformed the global baselines it analyzed across rain rates, including NASA's IMERG-Final precipitation estimates.

For light-to-moderate precipitation, the company says PARDIG improved hourly precipitation detection skill by more than 30% over IMERG-Final. The reported hourly detection evaluation against ground radar networks covered 2016 to 2023; DeepMind also says PARDIG was particularly effective at capturing extreme precipitation.

Both datasets' live and historical data are openly available on Google Cloud Storage. PARDIG's historical catalogs are also available in Google Earth Engine.

DeepMind wants to work with the wider community on models for floods, landslides and drought, particularly in regions without dense ground radar networks. The work involved teams at Google DeepMind and Google Research, in collaboration with Google Cloud and Google Geo.

Source

Companies in this story

Related stories

Explore

More articles