Google DeepMind opens AI rainfall dataset with records back to 1999
Google DeepMind releases global AI rainfall data
Google DeepMind announced on Nov. 13 that researchers and developers can now access PARDIG, its artificial intelligence system that estimates rainfall worldwide. The release includes historical records and continuously updated data.
PARDIG combines satellite observations and atmospheric data to estimate rainfall every 15 minutes on a global grid with roughly 10-kilometer resolution. Google said the system could help researchers track floods and droughts in regions with little weather infrastructure.
Key points about PARDIG
- PARDIG estimates rainfall every 15 minutes on a grid of roughly 10-kilometer cells.
- Historical records go back to 1999, with a separate dataset covering 2016 to the present.
- Google reports more than 30% higher detection skill than NASA's benchmark for light-to-moderate rainfall.
- Near-real-time data is published in six-hour batches, creating an average delay of about five hours.
- Both datasets are released under a Creative Commons license that allows commercial reuse with attribution.
What Google's announcement says
Google said PARDIG learns from precipitation measurements collected by NASA's Global Precipitation Measurement satellite. It also uses other satellite imagery and atmospheric information to estimate rainfall across the planet.
Unlike conventional systems that rely heavily on ground-based weather stations, PARDIG can produce estimates over oceans and areas with sparse observation networks. The dataset also supports Google's WeatherNext 3 forecasting model, which uses PARDIG rainfall information during training.
AI system outperforms NASA rainfall benchmark
DeepMind said PARDIG achieved more than 30% higher hourly precipitation detection skill than NASA's IMERG-Final dataset when measuring light-to-moderate rainfall. The company tested the system against radar networks in the United States, Europe and Japan, along with observations from more than 36,000 weather stations worldwide.
Google's research paper also reported improved identification of intense rainfall events, which existing global precipitation datasets can struggle to capture. The company highlighted the July 2025 Texas Hill Country floods as an example, comparing PARDIG's estimates with ground-based radar and NASA's existing dataset.
A five-hour delay limits emergency use
Despite the accuracy gains, the data arrives slowly. PARDIG produces estimates every 15 minutes, but Google publishes its near-real-time data every six hours in batches of 24 estimates. According to Google's documentation, the effective delay ranges from roughly two to eight hours, averaging about five hours.
That means emergency services cannot use the dataset as an immediate substitute for live weather warnings. Google also acknowledged that the system can underestimate snowfall in mountainous regions and is less reliable in certain polar areas.
Why the rainfall data matters
The release could expand access to rainfall information in developing countries and remote regions where limited radar coverage makes it hard to track storms, flooding and dry spells. The Creative Commons license may allow companies to use the data in agricultural monitoring, water management and disaster-risk applications.
Google still classifies PARDIG as experimental. The company says it does not replace official observations, forecasts or warnings, and users making public-safety decisions should rely on official meteorological agencies.
Sources
- CryptoSlate
- Google's PARDIG documentation