Newshot Google Open-Sources the Datasets Used by the WeatherNext 3 Weather Model: Including the SatWeave Satellite Imagery Mosaic Dataset and PARDIG Global Precipitation Information
IT之家 reported on November 13 that Google DeepMind today officially opened up SatWeave and PARDIG, the two foundational datasets underpinning the WeatherNext 3 weather model. The datasets provide real-time data streams and historical archives to researchers and developers worldwide, and can be used for modeling weather, hydrology, climate, and related downstream applications.
According to the introduction, SatWeave is a global imagery mosaic product integrating multiple geostationary weather satellites. It aggregates observation data from five satellites—GOES-East, GOES-West, Meteosat 0, the Meteosat Indian Ocean satellite, and Himawari—and addresses the challenge of differences among satellites in map projection, file formats, sensors, and scanning cadence. The dataset updates once per hour and provides a historical archive spanning more than ten years, enabling WeatherNext 3 to generate hourly global forecasts based on real-time cloud imagery and storm dynamics, whereas traditional analysis data typically refreshes only every six hours. Google also hopes the research community will extend it to scenarios such as solar power forecasting and wildfire monitoring.
PARDIG is a global precipitation reanalysis dataset based on an AI architecture. It uses a spherical graph-grid Transformer architecture similar to WeatherNext 3, directly takes the sparse spaceborne radar observations from NASA's Global Precipitation Measurement Core Observatory as the training target, and uses SatWeave imagery and atmospheric fields from the European Centre for Medium-Range Weather Forecasts as inputs. Its spatial resolution is about 10 km (0.1°), its temporal resolution is 15 minutes, and its average latency is about 5 hours.
Evaluations by Google against three major land-based radar networks in the United States, Europe, and Japan and more than 36,000 weather stations worldwide show that PARDIG outperforms mainstream baseline products such as NASA IMERG-Final across all rain intensity bins; its hourly detection skill for light to moderate precipitation improves by more than 30% compared with IMERG-Final, and its ability to capture extreme precipitation is particularly outstanding. Because the model learns directly from satellite observations, it also has reliable accuracy in regions lacking ground radar and rain gauges.
In addition, Google also opened the 1999–2026 PARDIG historical archive and the real-time data streams for SatWeave and PARDIG. The data are hosted on Google Cloud Storage and released under the CC BY 4.0 license.
Related articles
- Google DeepMind releases AlphaGenome 2 gene sequence model: ability to identify pathogenic variants nearly doubles
- Google completes talent acquisition of startup Mechanize, aiming to fill its own AI coding shortfall
- DeepMind engineer rebuts “Google can’t build frontier models”: Gemini 3.8 Cyber already outperforms Mythos in multiple tests
- Google AI responsibility team to be moved out of DeepMind; employees worry independence of safety evaluations will be compromised
- Report says Google DeepMind team no longer pursuing frontier model R&D work, may face large-scale layoffs
- Inside story of Google DeepMind leadership change: after Gemini was delayed by two months, its coding ability still lagged behind competitors; Brin personally supervised and demanded “full commitment”