Subscribe
0%
Illustration for: DeepMind opens SatWeave and PARDIG, the satellite and rainfall data behind WeatherNext 3: how to get it, limits

DeepMind opens SatWeave and PARDIG, the satellite and rainfall data behind WeatherNext 3: how to get it, limits

Google DeepMind and Google Research have released two experimental global datasets under CC BY 4.0, an hourly satellite mosaic and an AI rainfall record going back to 1999. They are free to read with a Google Cloud account, but the docs list caveats that matter for flood work and climate trends.

Listen to this article0:00 / 5:00

Google DeepMind and Google Research have made public two datasets that feed their WeatherNext 3 forecasting system: SatWeave, a single global mosaic of five geostationary weather satellites, and PARDIG, an AI-generated record of global rainfall every 15 minutes. Both are licensed CC BY 4.0 and stored as Zarr files in public Google Cloud Storage buckets, according to Google's developer documentation. DeepMind announced the release late last week. CASRAI dates the blog post to November 13, while StartupHub.ai reports it on November 15.

Why it matters: rain gauges and ground radar are sparse over oceans, mountains and much of the Global South, as CASRAI notes. A free, gap-filled global rainfall record at about 10 km resolution, with a live feed, could be useful to hydrologists, insurers and climate researchers. The documentation is also unusually frank about where the data is weak.

What's in the release

SatWeave PARDIG
What it is Merged imagery from GOES-East, GOES-West, Meteosat Prime, Meteosat Indian Ocean and Himawari AI estimate of rainfall rate, trained on NASA/JAXA GPM radar retrievals
Grid 0.1° (~10 km), global 0.1° (~10 km), global
Time step 30 min (2016–2025 archive); hourly from 2026 15 min
Live feed Hourly, ~1 h latency Near-real-time version, ~5 h average latency (2–8 h in practice), bucket updated every 6 h
History 2016 onward; all five satellites from Nov 14, 2018 Long-record version 1999–2023; near-real-time version 2016–present
Contents 17 bands: 7 reflectance, 10 brightness temperature One variable: precipitation rate in metres per hour
Where gs://satweave_1/ gs://pardig_1/ and two Earth Engine collections

The figures come from Google's SatWeave and PARDIG guides. Neither guide mentions a price. Both say anyone with a Google Cloud account can read the data without filling in an access request.

What DeepMind claims

DeepMind says it checked PARDIG against radar networks in the US, Europe and Japan and against more than 36,000 weather stations. As reported by StartupHub.ai and AlphaSignal, it claims PARDIG detects light-to-moderate rain more than 30% better hour by hour than NASA's IMERG-Final, with its biggest gains in heavy downpours. CASRAI adds that DeepMind also claims PARDIG beat MSWEP V3 and ERA5. These are DeepMind's own results. The method is described in an arXiv preprint (2611.13941), which the PARDIG guide asks users to cite.

StartupHub.ai says the hourly SatWeave feed lets WeatherNext 3 start a new global forecast every hour. Systems built on conventional analysis data wait six hours for each refresh.

The limits, from Google's own docs

  • Experimental, not official. Both guides say the data comes "as-is." It is not an official observation, forecast or warning, and it should not be anyone's only source for protecting life or property.
  • Five-hour lag. PARDIG's live feed is too slow for nowcasting. StartupHub.ai also notes that at 10 km it cannot match the detail of dense radar networks such as NOAA's MRMS.
  • Trend traps. The long-record and near-real-time versions of PARDIG use different inputs and satellites. Google warns that multi-decade trends spliced across the two may include shifts that have nothing to do with climate.
  • Known blind spots. Skill drops over high mountains. Snow and terrain-driven rain are underestimated in winter, and light snow and drizzle are hard to detect. Beyond about 67° latitude the values are extrapolated. Rates are capped at 75 mm/h, so the most intense storm cores can be clipped.
  • Satellite edges. SatWeave has no data beyond roughly ±60–65° latitude. It can also contain missing scan lines, mismatches where satellites overlap, sun glint and night-time glare in the visible bands. EUMETSAT, NOAA and JMA source data may carry their own terms.
  • Units. PARDIG stores an instantaneous rate in m/h, so multiply by 1,000 for mm/h. WeatherNext 3's experimental_tp_1hr field is an hourly total in metres, not a rate.

Who should look at it

The best fits are researchers and analysts in places with little radar, people training their own weather or flood models, and anyone who wants a global rainfall layer in Earth Engine. Teams that already have dense local radar should keep it, as StartupHub.ai notes. AlphaSignal advises checking PARDIG against local gauges before using it for flood or landslide warnings. The SatWeave guide includes Python examples using xarray and zarr, so trying a single region takes little effort.

Sources

  1. Google for Developers — WeatherNext: PARDIG guide developers.google.com
  2. Google for Developers — WeatherNext: SatWeave guide developers.google.com
  3. CASRAI — DeepMind opens datasets behind its AI weather forecasting models casrai.org
  4. StartupHub.ai — DeepMind opens live data behind its weather AI startuphub.ai
  5. AlphaSignal — Google DeepMind opens WeatherNext 3 data to boost rain detection by 30% alphasignal.ai

#Google DeepMind#Google Research#WeatherNext 3#SatWeave#PARDIG#weather#climate data#open data