Google DeepMind and Google Research released WeatherNext 3, a global weather forecasting model. It learns from live geostationary satellite images rather than only physics simulations. It produces a new forecast every hour at up to five-kilometre resolution. The earlier model worked on a coarser grid in six-hour steps. The model also trains on sparse weather station readings to capture local terrain. Google says this helps regions in Latin America, Africa and Asia-Pacific that lacked high-resolution forecasts. New variables cover turbine-height wind speed, cloud cover and solar radiation for renewable energy operators. The model starts powering weather in Search, the Gemini app, Google Maps, the Maps Platform Weather API and Earth Engine. Developers can query the data in BigQuery and Earth Engine or bulk-download it. Google cites independent live evaluations by Brightband for its accuracy ranking.
What changed
WeatherNext 2 forecast on a 25-kilometre grid in six-hour steps, trained on physics simulations.
What it unlocks
Querying hourly 5-kilometre global forecasts, including turbine-height wind and solar radiation, in BigQuery and Earth Engine.
- 5 km resolution for temperature
- hourly forecasts vs 6-hourly before
- up to 50% more accurate precipitation
- up to 60% CRPS gain vs IMERG
What you need to act on it
- Google Cloud access for BigQuery, Earth Engine or Cloud Storage data
- Google Maps Platform Weather API for developer use
Sources