Google DeepMind Opens WeatherNext 3 Data to Boost Rain Detection by 30%

Google DeepMind has released two observational datasets used by its WeatherNext 3 forecasting system. PARDIG and SatWeave are free to download from Google Cloud Storage, while PARDIG is also available through Google Earth Engine, according to the release announcement .

Forecast quality depends on timely, consistent observations. Satellite feeds differ in format, projection, sensor bands, and scan schedules, while direct rainfall measurements remain sparse across much of the world. SatWeave harmonizes global satellite imagery, and PARDIG turns satellite and atmospheric data into precipitation estimates. Both reduce the preprocessing required to build weather, hydrology, and climate tools.

Two datasets fill different gaps

SatWeave turns five feeds into one

SatWeave combines observations from GOES-East, GOES-West, Meteosat-0°, Meteosat Indian Ocean Data Coverage, and Himawari. These satellites cover different regions and use distinct projections, sensors, file formats, and scan schedules. Their observations can arrive 10 to 30 minutes apart, complicating efforts to assemble a synchronized global image.

DeepMind says many existing composites retain four or fewer channels or update only every few hours. SatWeave preserves 17 visible, water-vapor, and infrared channels in an hourly global product. The additional bands expose cloud properties, atmospheric moisture, surface temperature, smoke, and other features used by forecasting and Earth-observation models.

Hourly observations allow WeatherNext 3 to initialize forecasts from a newer global snapshot. Systems based on analysis data commonly initialize every six hours, leaving less recent conditions available when storms develop quickly. The live feed supports operational use, while the decade-plus archive provides training data for atmospheric vision models.

PARDIG fills radar gaps from orbit

PARDIG, short for Precipitation AI Reanalysis: Densely Inferred GPM, estimates rainfall in regions with limited ground-radar coverage. Its architecture is related to WeatherNext 3 and uses SatWeave imagery alongside atmospheric fields from the European Centre for Medium-Range Weather Forecasts.

Training targets come from the radar aboard NASA’s Global Precipitation Measurement Core Observatory. That instrument supplies direct but geographically sparse measurements as it orbits Earth. PARDIG learns the relationship between those measurements and the broader satellite and atmospheric inputs, then produces global precipitation maps at 15-minute intervals on a grid of roughly 10 kilometers.

The 15-minute interval describes the product’s temporal resolution, while its average delivery latency is about five hours. Developers building real-time alerts need to account for that distinction when setting ingestion schedules and warning thresholds.

Benchmarks show stronger detection

DeepMind evaluated PARDIG against NASA’s IMERG-Final, a widely used research product for global precipitation. The company reports an improvement of more than 30% in hourly detection skill for light-to-moderate precipitation. It also reports stronger detection of extreme rainfall, although the release does not attach a single summary percentage to that result.

The evaluation used continental radar networks in the United States, Europe, and Japan, plus observations from more than 36,000 weather stations. Those comparisons cover several climates and sensor networks, but teams should still validate performance against local gauges, terrain, and storm types before deploying high-consequence flood or landslide warnings.

Satellite-based inference extends precipitation coverage across areas with sparse radar infrastructure, including parts of Africa, South Asia, and South America. Local accuracy can still vary with terrain, climate, satellite viewing angle, and the availability of observations represented in the training and evaluation data.

Where developers can use the data

DeepMind presents both datasets as inputs for models and applications. Potential uses include:

• Flash-flood and landslide warning in regions without ground radar
• Drought monitoring and water-resource planning
• Wildfire-smoke tracking with geostationary imagery
• Solar-power nowcasting from cloud-cover channels
• Training Earth-observation models on long historical records
• Evaluating weather forecasts against recent precipitation estimates

The PARDIG archive spans 1999 to 2026, supporting studies of seasonal patterns and long-term model training. Its research paper describes the architecture and evaluation, while the WeatherNext developer docs cover access to live feeds.

Teams integrating the datasets should plan for several operational details:

• Separate cadence from latency. PARDIG represents 15-minute periods but arrives about five hours later on average.
• Validate regional performance. Global coverage does not guarantee uniform accuracy across climates and terrain.
• Inspect historical consistency. Satellite sensors, coverage, and upstream inputs can change across long archives.
• Check schemas and terms. Bucket paths, file formats, variables, and usage conditions should come from the current documentation.
• Budget for downstream logic. Forecasting, alert thresholds, uncertainty handling, and delivery systems remain application responsibilities.

A lower barrier to weather tooling

ERA5 provides a broad atmospheric reanalysis but commonly reaches users days after observation. IMERG-Final offers research-grade precipitation estimates with multi-month latency. PARDIG targets lower-latency precipitation, while SatWeave supplies hourly global imagery. Each product serves a different job, so selection should follow the variables, latency, and validation requirements of the application.

Publishing these datasets removes part of the satellite-ingestion and harmonization workload that previously required substantial infrastructure. Smaller teams can begin with consistent global imagery and precipitation estimates, then focus their engineering effort on forecasting, risk models, and user-facing systems. The release also makes a key part of WeatherNext 3’s observation pipeline available for independent research and comparison.