November 16, 2026:


Google DeepMind opened two datasets to the public on Friday, Nov. 13, and the headline claim is about rain: PARDIG, a global precipitation record at 0.1° resolution (about 10 km, or 6.2 miles), reaches users roughly five hours after observation, while NASA’s research-grade IMERG-Final product takes months. The lab’s own paper reports that PARDIG scores higher than both IMERG-Final and IMERG-Early across the thresholds and regions it tested.
Any Google Cloud account can download the data, with no access request form, under a Creative Commons Attribution 4.0 (CC BY 4.0) license. That gives flood modelers, hydrologists and weather developers a same-day global rain reference to train on and to score their own models against. The second release, SatWeave, is an hourly mosaic of geostationary satellite imagery with about one hour of latency, and it supplies the live observations that start WeatherNext 3 every hour.
Google labels both products experimental and as-is, and says PARDIG is not an official observation, forecast, watch or warning. The paper went to arXiv on Nov. 10, three days before the blog post.
NASA’s IMERG blends data from the GPM satellite constellation into half-hourly global rain estimates, and it comes in three runs. The Early Run arrives about four hours after observation, the Late Run follows after roughly 12 to 14 hours, and the Final Run waits for monthly rain-gauge analysis, which pushes its latency to about three months, with NASA’s pages giving figures from 2.5 to 3.5 months. NASA advises researchers to use the Final Run whenever latency allows.
PARDIG’s near-real-time feed lands in 24 frames every six hours. The newest frame is about two hours old, the oldest about 7.5 to 8 hours, and the average is about five hours. That puts PARDIG in the same latency class as IMERG-Early while, according to the authors, outscoring both Early and Final. DeepMind puts the gain over IMERG-Final at more than 30% in hourly detection skill for light-to-moderate rain.
The score behind those claims is the Critical Success Index (CSI), which rates how often rain is correctly detected and penalizes both misses and false alarms; 1.0 is perfect. Against more than 36,000 weather stations worldwide at 1 mm/h (0.04 in/h), the paper reports a mean CSI of 0.36 for PARDIG’s near-real-time version against 0.28 for IMERG-Final. Over the continental U.S. radar network, the scores are 0.48 against 0.38 at 1 mm/h and 0.37 against 0.26 at 4 mm/h (0.16 in/h). The authors say PARDIG posts the top score at every threshold in all three radar regions: the U.S., Europe and Japan.
The training signal comes from the radar and radiometer aboard the NASA and JAXA GPM Core Observatory. PARDIG trains on CORRA precipitation retrievals, from the Combined Radar-Radiometer Algorithm, and those retrievals are precise but cover only narrow swaths. The spacecraft’s dual-frequency radar sees a swath about 245 km (152 miles) wide at any moment. PARDIG learns to predict those sparse measurements from inputs that cover the whole planet continuously, then runs everywhere, including where the radar never looked.
The model is a spherical graph-mesh transformer built on an icosahedral mesh of 40,962 nodes, with 16 layers. Each run takes four hourly atmospheric frames from the European Centre for Medium-Range Weather Forecasts (ECMWF) and eight 30-minute geostationary frames spanning a four-hour window. Rather than predicting one number, it outputs a probability across 14 rain-rate bins, and one pass per hour produces global fields every 15 minutes. WeatherNext 3 uses a bigger version of this architecture with 32 layers.
The speed comes with a tradeoff. IMERG-Final’s delay exists because it waits for gauge analysis, while PARDIG does not use rain gauges as inputs; its listed inputs are atmospheric fields plus satellite imagery. Skill advantages shrink over high mountains: Google’s guide says skill margins narrow over mountains, and the paper reports IMERG-Final scoring higher at some stations in high-elevation terrain such as parts of the Rockies and the Alps. Jackson Tan, a researcher at NASA’s Goddard Space Flight Center, describes IMERG as weaker near orography, so terrain is a shared problem.
DeepMind’s blog describes a 1999-to-2026 record, but that span is two products joined together. The near-real-time version (PARDIG-NRT) runs from January 2016 to the present using ECMWF’s HRES forecasts and SatWeave, while the long-record version (PARDIG-LR) covers 1999 through 2023 as a static archive built from ERA5 reanalysis, NOAA’s GridSat-B1 satellite record and passive-microwave sounders.
The paper reports its largest station gains in the tropics. Between 30°S and 30°N, PARDIG-NRT averages a CSI of 0.36 against 0.24 for IMERG-Final, and it scores better at 96.8% of stations there. Poleward of 30°, the figures are 0.36 against 0.28.
IMERG-Final’s gauge adjustment is only as strong as the gauge network behind it. NASA says the gauge analysis plays an important or dominant role in well-gauged areas, while oceans mostly lack gauge input. A rain gauge also measures rain at a single point and can undercatch in strong wind. Whether gauge sparsity explains the tropical gap is a question for independent testers.
The ground truth behind the headline numbers is uneven. Radar validation relies on three independent radar networks in the U.S., Europe and Japan, and the authors write that dense radar networks are largely confined to a few developed regions. The 36,000-station network is the evidence for performance elsewhere.
SatWeave harmonizes imagery from five geostationary positions into one global grid: GOES-East and GOES-West from NOAA, Meteosat at 0° and over the Indian Ocean from EUMETSAT, and Japan’s Himawari. Its 17 bands comprise seven reflectance bands and ten brightness-temperature bands, and only 11 of them come from all five positions. Geostationary sensors see only about 60 to 65 degrees of latitude either side of the equator, so polar cells are empty.
The archive holds 30-minute steps from 2016 through 2025, and hourly data from 2026 onward, in cloud-native Zarr format. Frames can contain missing scan lines, calibration differences where satellites overlap and solar glint, and data from 2016 to mid-2017 is sparse over the Western Hemisphere.
DeepMind contrasts SatWeave’s hourly updates with six-hour analysis refreshes used by other models. That feed is what lets WeatherNext 3, which Google introduced on Sept. 3, produce hourly global forecasts at up to 5 km (3.1 miles) resolution, against 25 km (15.5 miles) for WeatherNext 2. PARDIG then serves as the precipitation training target. Google’s guide says that choice yields about 15% lower probabilistic rain-forecast error (CRPS) than training on IMERG, and about 30% lower error than a reanalysis-trained head and WeatherNext 2 at early lead times, measured against independent ground radar.
PARDIG’s deterministic rain rates are capped at 75 mm/h (3 in/h), which Google’s documentation says can clip peak convective rainfall. Estimates poleward of about 67° latitude extrapolate beyond the training domain, light snow and shallow drizzle are hard for the sensors to detect, and switching between the two variants can create non-climatic shifts, so multi-decade trend work needs care.
The hardest cases are the heavy ones. The July 2025 Texas Hill Country floods killed at least 135 people, most of them in Kerr County. For that county, the paper reports one-hour CSI scores of 0.29 for PARDIG against 0.08 for IMERG-Final at 15 mm/h (0.6 in/h), and 0.10 against 0.02 at 25 mm/h (1 in/h). PARDIG wins by a wide ratio, yet a CSI of 0.10 still means most heavy-rain detections are misses or false alarms. DeepMind’s blog compares PARDIG with NOAA’s roughly 1 km (0.6 mile) MRMS radar product and IMERG-Final for that event but publishes no numbers for it.
The benchmarks come from the dataset’s own creators in an arXiv preprint. Independent re-checks have challenged Google AI claims before: in 2026, hydrologists led by Kailong Li argued in Journal of Hydrology X that the real accuracy of Google’s AI flood forecasting system is likely well below what was reported, especially for extreme floods. That is a different product, but it shows why outside scoring of PARDIG matters.
A five-hour average delay cannot trigger a flash-flood warning. During the Texas event, the Guadalupe River rose 26 feet in 45 minutes, according to state officials. Google itself tells users to follow official national and local alerts for safety decisions.
The practical uses sit just behind the warning window: recalibrating hydrological models the same day, reconstructing storms in places with no radar, scoring nowcasting models against a common global reference, and training new rainfall models on the same inputs and targets that feed a deployed forecasting system. CC BY 4.0 allows commercial use if users credit the dataset.
Access depends on the product. Live and historical PARDIG data sit on Google Cloud Storage, while Earth Engine holds a static PARDIG-NRT archive through Oct. 22, so ongoing updates require Cloud Storage. The decision for a reader who works with rain data is a trade: PARDIG gives open, same-day, 15-minute global fields that beat NASA’s slowest product in the authors’ tests, while NASA’s Final Run still brings gauge anchoring, a long track record and independent scrutiny, and both products lose skill in mountains and at extremes.
CC BY 4.0 permits commercial use and adaptation as long as users give credit, which includes naming the creator, linking the license and stating any changes. Google still provides the data as experimental and as-is.
IMERG-Early arrives in about four hours and uses only forward propagation; NASA’s Version 06 applied no gauge adjustment to it. The paper says PARDIG scores higher than IMERG V07 Early as well as Final, with the largest gains at the highest intensities.
No. Its average delay is about five hours, Google says it is not an official observation or warning, and heavy-rain scores remain low in absolute terms. Local alerts from national weather services remain the source for safety decisions.
Not on its own. The 1999-to-2023 long-record version and the 2016-to-present near-real-time version use different inputs, and switching between them can introduce non-climatic shifts.