Google DeepMind opens weather satellite imagery and 15-minute precipitation estimates

Google DeepMind opened two datasets used in its AI weather research on Nov. 13 UTC. SatWeave brings imagery from several geostationary weather satellite positions onto a common grid. PARDIG uses satellite imagery and estimates of atmospheric conditions to calculate precipitation rates at 15-minute intervals. Researchers outside Google can use the data to train their own weather models or evaluate predictions against a historical record.

The release covers two datasets. DeepMind’s previously released WeatherNext 3 uses SatWeave imagery to initialize forecasts of future weather. PARDIG estimates precipitation for times that have already passed. A researcher looking back at a storm could use SatWeave to follow the movement of a cloud band and PARDIG to see where the estimated precipitation was concentrated around the same time.

Satellite imagery stitched across five orbital positions

SatWeave combines imagery from the two U.S. GOES positions, two European Meteosat positions and Japan’s Himawari. Each operator delivers images with its own map projections and instrument calibrations. Scan intervals range from 10 to 30 minutes. Google reconciles the measurements and blends overlapping views onto a latitude and longitude grid spaced 0.1 degrees apart. That gives a research team a common map when a weather system crosses from one satellite’s view into another’s.

SatWeave contains 17 spectral bands, including reflected visible light, water vapor and thermal infrared. Its 2016–2025 archive has images every 30 minutes. Records from 2026 onward are hourly, with new imagery arriving about an hour after observation. The grid extends to both poles, but the satellites see only to about 60 to 65 degrees north and south. The dataset marks cells beyond these limits as missing, along with bands a satellite does not carry and reflected sunlight at night.

Estimates every 15 minutes arrive in six-hour batches

DeepMind trained PARDIG using precipitation data from the radar and radiometer aboard the NASA–JAXA Global Precipitation Measurement Core Observatory. The satellite takes measurements along narrow strips as it orbits Earth. In its near-real-time version, PARDIG combines selected SatWeave bands with atmospheric fields from the European Centre for Medium-Range Weather Forecasts. It uses those inputs to estimate precipitation across a 0.1-degree grid, including locations outside the strips sampled by the training satellite.

PARDIG-NRT runs from 2016 onward. Google publishes 24 new frames on Cloud Storage every six hours. The newest frame in a batch is about two hours old when it appears, and the oldest is about eight hours old. The average delay is about five hours. The Earth Engine catalog describes the estimates as experimental data for research and modeling.

The research team’s preprint compared the estimates with ground radar across the contiguous United States, Europe and Japan for 2016–2023. In the U.S. comparison at a threshold of 1 millimeter per hour, PARDIG-NRT recorded a Critical Success Index of 0.48, against 0.38 for NASA’s IMERG Final. The index counts correct detections against the sum of correct detections, missed events and false alarms. The scores describe how well each dataset identified where precipitation had occurred at that threshold and in that region.

A separate archive reaches back to 1999

PARDIG-LR covers 1999 through 2023 with the same grid and 15-minute spacing. It draws on older satellite records and a historical reconstruction of atmospheric conditions from ECMWF. The two versions overlap from 2016 through 2023, allowing comparisons at the same times. Google’s documentation warns that changes to the observing system can also change the estimated readings. That matters if a study combines the records to measure how precipitation changed across decades.

Google’s documentation lists both datasets under CC BY 4.0 and says anyone with a Google Cloud account can access them on Cloud Storage without applying. The PARDIG-NRT copy in Earth Engine is a fixed archive ending on Oct. 22, 2026. The continuing feed for later dates is on Cloud Storage.