DeepMind opens SatWeave and PARDIG datasets behind WeatherNext 3: hourly 17-channel satellite mosaic and global precip maps with >30% detection-skill gain vs IMERG-Final

Why it matters

【Key points】

· Release: DeepMind’s WeatherNext team posted at 12:00 UTC on 13 Nov 2026, opening SatWeave and PARDIG, two datasets that fed September’s WeatherNext 3.

• · SatWeave: hourly 17-channel global geostationary mosaic (visible, water-vapor, infrared), with a live feed and a 10+ year archive.
• The body names five satellites (GOES-East, GOES-West, Meteosat 0, Meteosat Indian Ocean, Himawari); a caption says “six” but does not name a sixth.

· PARDIG: ~15-minute / ~10 km resolution, ~5-hour average latency. Trained on sparse NASA GPM radar, with SatWeave imagery and ECMWF fields as inputs.

· Evals: US, European and Japanese radar networks plus 36,000+ stations. DeepMind says it beats NASA IMERG-Final and lifts hourly precipitation detection skill (CSI, 2016–2023) by more than 30% on light-to-moderate rain.

· Access: PARDIG archives cover 1999–2026, on Google Cloud Storage; catalogs also in Google Earth Engine. No bucket or asset IDs are given.

【Our read】

Background

• observational gaps often limit weather models as much as the forecast network.
• WeatherNext 3 launched in September without these two datasets.
• This post opens that data layer.

Comparison

• Most global satellite composites keep four or fewer channels, or update every few hours.
• The “more than 30%” CSI figure is DeepMind’s own 2016–2023 evaluation, not an independent replication, and no full regional table is published.
• This is not a new forecast model, and WeatherNext 3 weights are not being opened.

Impact

live feeds plus long archives lower the barrier for flood, drought, landslide, solar and wildfire applications.

【Verifiable】

• · DeepMind blog “Opening the datasets behind our AI weather models” (13 Nov 2026, datePublished 2026-11-13T12:00:00Z)
• · WeatherNext 3 (September 2026)