SI Agent News
DevelopingGeminiResearch

Google DeepMind opens near-real-time global rain dataset

The new PARDIG dataset provides global precipitation data with a five-hour latency, significantly faster than NASA's research-grade records.

0.36PARDIG mean CSI against 36,000 weather stations

The short version

Why it matters

This provides hydrologists and weather developers with a high-quality, same-day global reference for training and scoring models. It bridges the gap between slow, high-accuracy historical records and fast, less-accurate real-time estimates.

Closing the latency gap

Google DeepMind has released two new datasets, PARDIG and SatWeave, designed to provide global rain and satellite imagery. PARDIG offers a global precipitation record at a 0.1° resolution, which translates to roughly 10 km or 6.2 miles. The primary goal of this release is to provide users with a near-real-time reference that is significantly faster than current research-grade standards.

The headline claim involves a major speed advantage over NASA's IMERG-Final product. While NASA's Final Run takes about three months to process because it waits for monthly rain-gauge analysis, PARDIG reaches users in about five hours. This puts the new dataset in the same latency class as NASA's IMERG-Early product, but the authors claim it achieves higher accuracy scores.

As Tech Times reported, the lab's own paper indicates that PARDIG performs better than both IMERG-Final and IMERG-Early across various regions and thresholds. This means flood modelers and weather developers can now access same-day global rain data to train their own models. The data is available to anyone with a Google Cloud account under a Creative Commons Attribution 4.0 license.

How the model processes rain

PARDIG estimates rain in areas where there is no radar by learning from sparse measurements. It uses training signals from the radar and radiometer systems on the NASA and JAXA GPM Core Observatory. Specifically, it trains on CORRA precipitation retrievals, which are precise but only cover narrow swaths of the planet at any given time.

The model is a spherical graph-mesh transformer built on an icosahedral mesh of 40,962 nodes with 16 layers. It uses four atmospheric frames every hour from the European Centre for Medium-Range Weather Forecasts and eight 30-minute geostationary frames spanning a four-hour window. Instead of predicting a single number, the model outputs a probability across 14 different rain-rate bins.

This architecture allows the model to run everywhere, including regions the radar never visits. One pass per hour produces global fields every 15 minutes. While the speed is a major benefit, there are trade-offs; PARDIG does not use rain gauges as inputs, which means it lacks the specific ground-truth anchoring that NASA's slower products provide.

Performance and limitations

The performance of PARDIG was measured using the Critical Success Index, which rates how often rain is correctly detected while penalizing misses and false alarms. In tests against more than 36,000 weather stations worldwide, PARDIG's near-real-time version achieved a mean score of 0.36. This outperformed the 0.28 score recorded by NASA's IMERG-Final product for light-to-moderate rain.

The largest gains were observed in the tropics, where PARDIG scored better at 96.8% of stations between 30°S and 30°N. However, the model's skill margins narrow in high-elevation terrain like the Rockies and the Alps. In these mountainous regions, NASA's IMERG-Final sometimes scores higher because of the complexities of mountain weather.

There are also physical limits to the data. PARDIG's deterministic rain rates are capped at 75 mm/h, which can clip the peak of heavy convective rainfall. Additionally, the model struggles with light snow and shallow drizzle because these are difficult for the sensors to detect. Because of these factors, Google notes that the data is experimental and should not be used for official weather watches or warnings.

Practical applications for users

Because of the five-hour average delay, PARDIG cannot be used to trigger flash-flood warnings. For example, during a 2025 flood in Texas, a river rose 26 feet in just 45 minutes, a timeframe much faster than the dataset's update cycle. Google explicitly advises users to follow official national and local alerts for safety decisions during such events.

The practical utility of the dataset lies in the window just behind real-time warnings. Hydrologists can use the data to recalibrate hydrological models on the same day a storm occurs. It also allows researchers to reconstruct storms in regions that lack radar infrastructure or to score their own nowcasting models against a common global reference.

Developers can also use PARDIG as a training target for new rainfall models. By using the same inputs and targets that feed a deployed forecasting system, they can improve the accuracy of their own tools. The Creative Commons license allows for commercial use, provided that users give proper credit to the creators and state any changes made to the data.

The role of SatWeave

The second release, SatWeave, serves as a foundational component for these weather tools. It creates an hourly mosaic of geostationary satellite imagery from five different positions, including sensors from NOAA, EUMETSAT, and Japan. It harmonizes 17 different bands into a single global grid, providing the live observations that start Google's WeatherNext 3 every hour.

SatWeave contains 30-minute data points from 2016 through 2025 and hourly records starting in 2026. The imagery covers about 60 to 65 degrees of latitude on either side of the equator, meaning polar cells remain empty. The data is provided in a cloud-native Zarr format, which is a standard for storing large arrays of data in the cloud.

By providing hourly updates, SatWeave allows for more frequent analysis than the six-hour refreshes used by some other models. This high-frequency feed is what enables WeatherNext 3 to produce global forecasts at a 5 km resolution. Together, PARDIG and SatWeave provide a comprehensive suite of tools for the global weather research community.

What to watch

Try it

If you have a Google Cloud account, you can download the PARDIG dataset today to test its rain-rate bins against your own local weather data.

Sources and links

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

Listen to the AI Wire headlines

Listen to the AI Wire

Listen to our daily AI shows every morning: on your commute, in the shower, wherever you are

Listen to the daily shows

First reported by Tech Times · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

More AI news