News
DeepMind Opens the Data Behind WeatherNext 3: SatWeave Satellite Mosaic and PARDIG Rainfall Maps
Google DeepMind has released two datasets, SatWeave and PARDIG, that underpin its WeatherNext 3 forecasting model. According to the company, PARDIG improves hourly detection of light-to-moderate precipitation by more than 30% over IMERG-Final.
On November 13, 2026, Google DeepMind published two datasets that power its WeatherNext 3 weather forecasting model. SatWeave is a global mosaic of imagery from geostationary satellites, and PARDIG is a set of precipitation maps at 15-minute resolution.
Two datasets
SatWeave combines imagery from several geostationary satellites: GOES-East, GOES-West, Meteosat 0, Meteosat Indian Ocean and Himawari. The source scans run on different schedules, from 10 to 30 minutes, and use different projections, formats and sensors, so DeepMind merged them into a single global mosaic. The DeepMind blog post is inconsistent on this point: a video caption refers to six satellites, while the text lists five.
According to DeepMind, existing global composites typically offer four or fewer channels or update only every few hours. SatWeave has 17 channels and updates hourly. Both a live feed and a historical archive are available.
PARDIG, short for Precipitation AI Reanalysis: Densely Inferred GPM, uses an AI architecture similar to WeatherNext 3. The model was trained on sparse observations from NASA's GPM Core Observatory spaceborne radar, and its inputs are SatWeave imagery and atmospheric fields from ECMWF. The output is a global precipitation estimate every 15 minutes at a resolution of about 10 km.
What it means for WeatherNext 3
SatWeave's hourly refresh lets WeatherNext 3 start new global forecasts more often. Models initialized on analysis data refreshed every six hours cannot do that.
DeepMind says it is releasing the data to help researchers and developers build weather, hydrology and climate applications. The company also points to other uses: solar power forecasting and wildfire monitoring for SatWeave, and flood, landslide and drought models for PARDIG, especially where dense ground radar networks are missing.
Results and caveats
PARDIG was evaluated against radar networks in the US, Europe and Japan (2016-2023) and more than 36,000 weather stations worldwide. According to DeepMind, the model outperforms leading global baselines across all analyzed rain rates, including IMERG-Final. As an example, the company shows the July 2025 flash floods in Texas, compared with NOAA MRMS radar.
There are caveats. A resolution of about 10 km is coarser than ground radar (about 1 km), and the roughly five-hour latency is an average. Reliable accuracy in regions without ground measurements is an expectation on DeepMind's part, not a demonstrated result. The post also does not state the data's licensing terms.
Who it is for
The datasets are aimed at research teams and developers building weather, hydrology and climate applications. The DeepMind post links to the WeatherNext developer documentation and to a research paper describing PARDIG. For Polish teams working on precipitation forecasting or solar energy, it is an additional data source to test, though the licensing terms should be checked before any commercial use.
