DeepMind opens AI‑driven weather datasets, giving researchers hourly satellite views and high‑resolution rain maps

Farmers, renewable‑energy operators, and emergency managers all need weather information that is both timely and detailed, but many public models still deliver updates only every six hours. DeepMind’s latest announcement changes that landscape by making two core datasets—SatWeave and PARDIG—available to anyone, turning raw satellite observations into ready‑to‑use inputs for weather, hydrology and climate work.

SatWeave is an hourly, global mosaic of geostationary satellite imagery, while PARDIG is an AI‑derived precipitation product that delivers a global rain map every 15 minutes at roughly 10 km resolution. Both live feeds and a decade‑long historical archive are hosted on Google Cloud, giving developers the same data that powered DeepMind’s WeatherNext 3 model without building the preprocessing pipeline themselves.

Source: Google DeepMind Blog: Opening the datasets behind our AI weather models ↗

In a DeepMind blog post the team confirmed that the SatWeave and PARDIG buckets are openly hosted on Google Cloud Storage, and that the precipitation archive is also accessible through Google Earth Engine. The post notes that these datasets were central to the performance gains reported for WeatherNext 3, DeepMind’s most advanced global weather AI model to date.

SatWeave stitches together imagery from five geostationary satellites—GOES‑East, GOES‑West, Meteosat‑0, Meteosat‑Indian Ocean and Himawari—and refreshes the composite every hour, far more frequently than the six‑hour analysis cycles used by many operational systems. PARDIG is trained on sparse space‑borne radar from the NASA GPM mission together with SatWeave and ECMWF atmospheric fields, and it improves precipitation detection skill by more than 30 % over NASA’s IMERG‑Final for light‑to‑moderate rain while also capturing extreme events better.

Source: Google DeepMind Blog: Opening the datasets behind our AI weather models ↗

The datasets are free to download, but they are large and require cloud‑based storage and compute resources to process, so users must be comfortable with Google Cloud or Earth Engine environments. PARDIG’s precipitation maps have an average latency of about five hours, which can limit their usefulness for ultra‑rapid decision‑making such as flash‑flood warnings.

Data quality also depends on the underlying satellite observations; regions with persistent cloud cover or limited satellite view may receive less reliable inputs. Because the datasets have only just been released, community‑built validation tools and benchmark studies are still emerging, meaning early adopters may need to invest extra effort to verify results for their specific applications.

Researchers and developers can start by visiting the DeepMind blog links to the SatWeave and PARDIG buckets on Google Cloud Storage. The accompanying documentation explains how to stream the hourly feeds, query the historical archives, and ingest the data into existing modeling pipelines. For quick prototyping, load the PARDIG archive into Google Earth Engine and overlay it with local gauge networks to assess regional accuracy.

With the data in hand, you can build downstream services such as flood‑risk alerts, solar‑power output forecasts, or wildfire‑propagation models that benefit from the higher temporal resolution. Keep an eye on community forums and forthcoming research papers that benchmark these datasets, and consider contributing your own validation results to help the ecosystem mature.