DeepMind opens WeatherNext 3 training datasets

Google DeepMind has announced the public release of two Earth-observation datasets, SatWeave and PARDIG. The company says both were key to the gains in its recently released WeatherNext 3 forecasting model.

SatWeave is a global mosaic of geostationary satellite imagery. In WeatherNext 3 it gives the model a live view of clouds and storms as they develop. It combines data from five satellites that each cover a different part of the globe: GOES-East, GOES-West, Meteosat 0, Meteosat Indian Ocean and Himawari. These instruments differ in map projection, file format, sensor and scan schedule, with scans every 10 to 30 minutes. That makes a coherent global picture hard to build. DeepMind says most existing global composites deal with the mismatch by dropping data, so they offer four or fewer channels or update only every few hours.

PARDIG is an AI-based global precipitation dataset. DeepMind used it to train WeatherNext 3's precipitation forecasts.

DeepMind is releasing both historical archives and live feeds for each dataset, so they can be used in modeling work and downstream applications. The company presents the release as a way to tackle a core bottleneck in forecasting: turning raw satellite and ground measurements into a coherent global picture quickly enough to be useful.

The available excerpt leaves several things unstated: the licensing terms, how to access the data, how many channels SatWeave has or what resolution it offers, and how PARDIG is built or validated.