Google launches WeatherNext 3, hourly forecasts up to 5 km

Google introduces WeatherNext 3, an AI model that relies on live satellite observations to produce forecasts updated every hour. The company reports measured gains in rainfall prediction and immediate availability in Search, Maps, Gemini, and cloud APIs.

Precipitation: Gains Announced Up to 60% According to Metrics

In the assessment of rainfall in medium-term global forecasting, Google reports an improvement in the CRPS score of up to 60% compared to IMERG, 30% for MRMS, and 10% against rainfall measurements for early forecast timelines. The company also states that for planning a day or more ahead, users will see precipitation forecasts that are up to 50% more accurate. The most significant improvements are noted in areas where forecasts have historically been less reliable.

Data Sources: Live Geostationary Satellites at the Core of the Model

WeatherNext 3 ingests a mosaic of live geostationary satellite data to obtain a continuously updated view of the atmosphere. This choice allows for the generation of a new forecast every hour, based on the most recent observations, with a resolution of up to 5 kilometers. In contrast, many AI models, including the previous generation, are trained on outputs from numerical weather prediction models that are six hours behind and can introduce biases for rapidly changing variables like rainfall or surface temperature. To enhance rainfall prediction, the training combines NASA's IMERG data with a global precipitation reanalysis derived from satellite radar conducted by Google. The company emphasizes that rainfall forecasting is a challenging point for global models, with AI approaches often producing vague estimates or missing the boundaries of storms.

Access: Search, Maps, Gemini, and APIs Begin to Be Powered

Google indicates that WeatherNext 3 starts today to power weather experiences in Google Search, the Gemini app, Google Maps, the Weather API of Google Maps Platform, and Google Earth Engine. High-resolution global forecasts, updated every hour, are offered and ready to be integrated into workflows without model configuration. Enhanced forecasts are also available via Google Search, Google Maps, and Gemini, and can be integrated through Google Cloud. Operational integration is underway within the company's ecosystem, with a stated goal of utility for emergency interventions, air traffic control, and agriculture. Google highlights expected benefits for agriculture, renewable energy, and daily planning.

Resolution: From 5 to 25 km, and a Leap Compared to WeatherNext 2

WeatherNext 3 generates temperature and humidity maps with a resolution of 5 kilometers, offers other surface variables at a scale of 10 kilometers, and provides atmospheric variables such as wind at 25 kilometers, while ensuring physical consistency between the global scale and local topography. The forecasts are updated every hour and, according to Google, they are now five times more accurate than before. Compared to WeatherNext 2, which operated on a 25-kilometer grid with outputs every 6 hours, this represents a leap in resolution and frequency. The new model, launched by Google, also benefits from the fact that each new hourly forecast relies on the most recent satellite observations.