Google Opens Satellite and Rainfall Data for Weather and Flood Research

A flood model can misjudge a river’s rise when it misses rain upstream. In regions with few rain gauges or patchy radar coverage, establishing where the water fell is a central problem for researchers trying to anticipate where it will go.

Google opened two datasets on November 13 that could help fill those gaps. SatWeave brings satellite views of clouds and the atmosphere into a common format. PARDIG uses artificial intelligence to estimate rainfall across the planet, including places where observations from the ground are sparse.

The release gives other teams access to historical records and continuing data feeds that Google used in WeatherNext 3. Developers could use them to train forecasting systems, study past storms or build tools for managing water and renewable power.

WeatherNext 3 arrived in September. The new release opens some of the material behind it: SatWeave supplies observations used as forecast inputs, while PARDIG supplied rainfall estimates for training. The opportunity extends to researchers building their own models and applications.

Combining the satellite views

Weather satellites observe different regions using different instruments, file formats and scanning schedules. Google’s WeatherNext team describes the challenge as forming “a coherent global picture.” SatWeave organizes those observations into a shared map, reducing a substantial preparation job for teams that want to work with them.

The dataset includes 17 channels, each capturing a different part of the light spectrum. Visible-light channels show reflected sunlight; infrared measurements provide information about temperature and water vapor. Together they offer several ways to examine the same developing weather system.

SatWeave’s grid spacing is about 10 kilometers. Its archive starts in 2016, with half-hourly records through 2025 and hourly records from 2026 onward. Coverage varies by channel, location and satellite generation, with gaps toward the poles.

Fifteen minute detail with a five hour delay

PARDIG learns from precipitation measurements collected by the Global Precipitation Measurement satellite mission, led by NASA and the Japan Aerospace Exploration Agency. It combines satellite imagery with atmospheric information from the European Centre for Medium-Range Weather Forecasts to estimate rainfall between the satellite’s relatively narrow measurement paths. Its output is an estimated rain rate on a roughly 10-kilometer grid.

The maps represent conditions at 15-minute intervals. Delivery follows a different clock: new PARDIG estimates arrive in six-hour batches, averaging about five hours behind the conditions they describe. Depending on their position in a batch, individual maps are roughly two to eight hours old when released.

The delayed record can help reconstruct a storm’s progression and test a flood model against past events. Fast-moving emergencies require information that reaches decision-makers in time. SatWeave’s live feed has its own schedule, updating hourly with a target delay of about one hour.

PARDIG also comes in two historical versions. The near-real-time series covers 2016 onward; a companion archive reaches back to 1999 and ends in 2023. Their overlapping years allow comparisons, while changes in instruments and processing need attention when studying long-term trends.

What the results establish

In a preprint submitted November 10, Google researchers evaluated PARDIG against radar networks in the United States, Europe and Japan, and more than 36,000 weather stations worldwide. They reported stronger precipitation detection than several existing global datasets, including NASA’s IMERG products.

Google highlights an improvement of more than 30 percent over IMERG-Final in a measure of hourly light-to-moderate rain detection. That measure, the Critical Success Index, rewards correctly detected rain while penalizing missed events and false alarms. It evaluates whether rainfall exceeds a chosen intensity threshold.

Using an earlier internal version of PARDIG, the researchers trained one part of WeatherNext 3 to predict rainfall. It performed better against independent radar observations than the part trained on IMERG-Final. Those results support its value as training material. The open release now enables outside teams to test how those gains carry into their own regions and applications.

Available for reuse

Both the live feeds and historical archives of PARDIG and SatWeave carry the Creative Commons Attribution 4.0 license. It permits sharing and adaptation, including commercial use, with appropriate credit, a license link and disclosure of changes. Google also identifies third-party attribution and terms that users need to retain.

The data is available through Google Cloud Storage, with historical PARDIG collections also available in Earth Engine. The observational datasets are accessible without applying for the separate WeatherNext forecast-data allowlist.

For a team studying a poorly monitored watershed, the immediate gain is a shared record it can inspect and test. Turning that record into dependable local decisions will require checking its errors against the terrain, storms and observations that matter there.

Sources

• Google DeepMind announcement, November 13

• WeatherNext 3 introduction, September 3

• SatWeave dataset documentation

• PARDIG dataset documentation

• Data dissemination schedule

• PARDIG research preprint

• Dataset access guide

• Google dataset terms and attributions

• Creative Commons Attribution 4.0