Google opens PARDIG and SatWeave weather datasets
Google DeepMind announced open access to PARDIG and SatWeave on November 13, releasing precipitation estimates and harmonized satellite observations used in WeatherNext 3. For developers, the important distinction is access: these observational datasets do not require the application used for Google’s operational forecast feeds.
The announcement covers historical archives and continuing feeds, not a new consumer weather app. Google’s access guide identifies both as publicly accessible under CC BY 4.0. A Google Cloud account is still part of the documented workflow.
Choose imagery, precipitation estimates or forecasts
SatWeave supplies satellite imagery; PARDIG estimates rainfall and other precipitation from observations and atmospheric inputs. WeatherNext 3 produces predictions. These are related components, not interchangeable products. Opening the inputs lets researchers build their own analyses rather than only consume Google’s forecast outputs.
The SatWeave guide describes 17 spectral bands on a 0.1 degree grid, roughly 10 kilometers. Its 2016 to 2025 archive has half-hourly timesteps. From 2026, the archive and live stream use hourly timesteps, with approximately one hour of latency. Data is supplied as Zarr stores in gs://satweave_1/.
Coverage needs care. A global grid does not mean every channel observes every location: polar regions, unavailable sensor bands and nighttime reflectance can contain missing values. Historical satellite changes also affect coverage. The documentation describes five operational satellite positions, rather than a single uniform sensor.
PARDIG’s 15-minute resolution is not a 15-minute delivery promise
Google’s PARDIG documentation separates two variants. PARDIG-NRT covers 2016 onward, while PARDIG-LR is a static 1999 to 2023 record. Both provide precipitation estimates at 15-minute intervals on a roughly 10-kilometer grid.
The delivery schedule shows that NRT arrives on Cloud Storage in six-hour batches. Effective latency ranges from roughly two to eight hours, averaging about five. Earth Engine hosts static archives at launch, not that continuing live stream. Choose the Cloud Storage NRT path for updates and the Earth Engine collections for historical geospatial analysis.
The Earth Engine catalog specifies that the precipitation variable is an instantaneous rate in meters per hour, not accumulated rainfall. Multiply by 1,000 for millimeters per hour. That distinction matters before combining these observations with a forecast variable that represents an hourly total.
What the research establishes, and what it does not
The PARDIG preprint, submitted November 10, describes a graph-mesh transformer trained against sparse spaceborne radar measurements. Its authors report comparisons with radar networks in the US, Europe and Japan and more than 36,000 weather stations, with improvements over several global precipitation baselines. Those are the researchers’ evaluations, not ByteForward testing or a guarantee for every watershed.
Google flags limitations in polar extrapolation, complex terrain, snowfall and light drizzle. Changing inputs across the long record can introduce non-climatic shifts, so stitching LR and NRT together is not automatically a clean climate-trend series. Localized extremes can also encounter the model’s precipitation-rate cap.
Open data still needs attribution and validation
CC BY 4.0 permits reuse with attribution; it does not make an experimental estimate an official warning. The Earth Engine catalog says the data is for experimental modeling and is not validated or approved for real-world use. Follow the attribution and provider acknowledgements in the dataset guides, and validate coverage, units and latency for the intended application. Do not use PARDIG alone for decisions protecting life or property: Google’s guidance directs users to official meteorological and emergency alerts.
The practical starting point is a small historical slice, then a check of missing values and timing before adding a live pipeline. The release removes an access-request step for these datasets. It does not remove the engineering work needed to establish whether they fit a particular monitoring or research task.
Featured image: AI-generated editorial illustration, not an observational weather map. Google and the Google logo are trademarks of Google LLC. ByteForward is not affiliated with or endorsed by Google.
