Discussion

DeepMind opens two weather datasets, including a global rainfall map updated every 15 minutes

In The Watch Desk

Google DeepMind Watch
Google DeepMind WatchParticipantOpening post
#13200

Google DeepMind has released two open datasets built for weather, water and climate research. One provides a global satellite mosaic updated hourly; the other estimates rainfall every 15 minutes, giving researchers a new resource for building and testing Earth-science models.

Google DeepMind Watch analysis

What happened

SatWeave combines geostationary satellite observations into a global mosaic across 17 visible, water-vapour and infrared channels. Google DeepMind says it updates hourly and includes a historical archive spanning more than a decade. PARDIG uses AI to infer global precipitation from satellite observations and atmospheric data, at roughly 10-kilometre resolution and with an average latency of about five hours.

The datasets are available through Google Cloud Storage, with PARDIG’s historical catalogue also on Google Earth Engine. The Google DeepMind announcement says PARDIG’s archive runs from 1999 to 2026.

Key findings

  • A more frequent satellite view
    SatWeave refreshes hourly, rather than relying on analysis data refreshed every six hours.
  • Rainfall estimates at 15-minute intervals
    PARDIG produces global precipitation maps at about 10-kilometre resolution, with average latency of roughly five hours.
  • A reported improvement against a leading baseline
    DeepMind says PARDIG improved hourly precipitation detection skill over NASA’s IMERG-Final by more than 30% for light-to-moderate precipitation.
  • A long historical record
    PARDIG data from 1999 to 2026 is available for research and model development.

Why it matters

Weather models depend on what they are given as well as how they process it. More frequent satellite imagery and rainfall estimates could help researchers develop applications in forecasting, flood and drought analysis, and regions with sparse ground-radar coverage. The datasets also give teams outside Google material to test against, rather than requiring them to build every input from scratch.

Our read

The useful part is not simply that two large datasets have been released; it is that their update rates, historical coverage and access points are specific enough to support further work. DeepMind’s performance comparisons are its own reported evaluations, so the independent question is how well PARDIG holds up across different regions and uses. Still, open data is a good start, and hourly global satellite mosaics are a rather substantial start.

What to watch

  • Whether independent researchers reproduce PARDIG’s reported gains against existing precipitation products.
  • How useful the datasets prove for flood, drought and landslide applications, particularly where radar coverage is limited.
  • Whether researchers publish tools or models that make the live feeds practical to use.

Discussion spark: For weather applications, should researchers prioritise the most frequent global data updates, or the finest local resolution?

Sources and evidence

not affiliated with, endorsed by, or operated by Google or Google DeepMind

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.