What happened
Google DeepMind announced the release of SatWeave and PARDIG on November 13, 2026. The company describes both as foundational datasets behind WeatherNext 3 and says their live feeds and historical archives are now available to researchers and developers. The datasets are intended to support applications in weather, hydrology and climate, including work on solar power, wildfire monitoring, floods, landslides and drought.
SatWeave is a global mosaic of geostationary satellite imagery. It combines observations from six satellites across 17 visible, water-vapor and infrared channels, and updates hourly. Google DeepMind says the underlying satellites use different map projections, file formats, sensors and scan schedules, ranging from 10 to 30 minutes, which makes a unified global view difficult to produce. The company says SatWeave’s hourly cadence allows WeatherNext 3 to generate new global predictions, while models initialized on analysis data refreshed every six hours have a slower starting-data cycle. The release includes a historical archive spanning more than a decade.
PARDIG, short for Precipitation AI Reanalysis: Densely Inferred GPM, is an AI-based global precipitation dataset. Google DeepMind says it is trained directly on sparse spaceborne radar observations from NASA’s Global Precipitation Measurement Core Observatory, using SatWeave imagery and atmospheric fields from ECMWF as inputs. It produces precipitation maps at 15-minute temporal resolution and roughly 10-kilometer spatial resolution, with an average latency of about five hours. Its historical archive covers 1999 to 2026.
The company reports evaluating PARDIG against three continental radar networks in the United States, Europe and Japan, as well as more than 36,000 weather stations worldwide. It says PARDIG outperformed leading global baselines across analyzed rain rates, including NASA’s IMERG-Final, and improved hourly detection skill for light-to-moderate precipitation by more than 30%. Google DeepMind says it expects the dataset to be useful in areas without dense ground radar coverage. Both datasets are available on Google Cloud Storage; PARDIG’s historical catalogs are also available in Google Earth Engine.
What the archive adds
New Runtime’s earlier coverage on October 21, 2026 described Google Research’s geospatial reasoning agent combining Gemini with Earth AI foundation models (earlier coverage). This release is a new event in the same broader Earth AI work: it opens foundational observation data, rather than announcing another agent or model. New Runtime also covered Ai2 and Google’s climate-model collaboration on November 4, 2026 (earlier coverage); that initiative focuses on coupled atmosphere-ocean simulations, distinct from these datasets.
Trend signal
This release supports the idea that scientific AI needs inspectable data artifacts alongside models: researchers can access historical and live observations to build and evaluate downstream applications. It is adjacent to, but not a direct instance of, the evidence-package principle in Scientific agents need artifact evidence. The release opens data infrastructure, not a scientific agent workflow or a complete rerunnable evidence package.
Why now
Google DeepMind says SatWeave and PARDIG were used in WeatherNext 3, announced in September 2026. Opening the data now extends access to components used in that model and invites outside researchers to explore applications beyond medium-range forecasting.
Why it matters
Weather models are only as useful as the observations that describe current conditions and the historical records used to train and assess them. A shared hourly satellite mosaic and a precipitation product with a 15-minute time step can give research teams more frequent inputs than a six-hourly analysis cycle, while the archives offer material for retrospective testing. That could help teams investigate local forecasting problems such as flood or drought risk, especially where radar coverage is sparse.
The release also makes an important distinction visible: data can be useful independently of the model that helped produce or consume it. Teams can explore PARDIG as an input or comparison dataset for their own methods, and SatWeave as an observation layer for applications such as solar forecasting or wildfire monitoring. Google’s reported evaluation results are promising, but builders should treat them as claims to test against their own regions, thresholds and operational requirements rather than as proof of universal performance.
Limits and uncertainty
The reported improvement over IMERG-Final is Google DeepMind’s evaluation claim; the announcement does not establish that the same gain will hold for every location, precipitation threshold or downstream task. PARDIG’s roughly 10-kilometer resolution and average latency of about five hours may not suit applications that require finer spatial detail or lower latency. The company expects accuracy in regions without dense ground measurements, but the announcement does not independently verify that expectation. Availability is stated through Google Cloud Storage and Earth Engine; access terms, quotas and service-level commitments are not specified here.
Skill to build now
If you work on weather or hydrology models, take one local forecasting task and compare its current input data with PARDIG’s 15-minute, roughly 10-kilometer precipitation maps. Define the geographic area, time window and detection thresholds first, then measure where the new data changes errors or lead-time decisions. This is a practical way to test whether a global product helps your region before incorporating it into a production pipeline.
Reader takeaway
Google DeepMind has opened two datasets behind WeatherNext 3, not just published model results: SatWeave supplies hourly multi-satellite imagery, while PARDIG provides frequent global precipitation estimates and a long archive. Their practical value will depend on how well they hold up for a team’s specific region and task.