Google introduced WeatherNext 3: weather forecasting learns from satellite images
Google DeepMind and Google Research announced WeatherNext 3 — their most advanced global weather forecasting model. The main difference from predecessors: it trains directly on satellite observations rather than results of traditional meteorological calculations. This allows the model to respond faster to atmospheric changes and provide more localized forecasts.
WeatherNext 3 provides hourly forecasts with spatial resolution down to 5 kilometers. In comparison, the previous version WeatherNext 2 operated on a 25 km grid and updated every 6 hours. According to Google, the new model gives approximately five times more detailed weather picture.
The greatest progress has been achieved in precipitation accuracy. According to estimates based on several datasets, the improvement in the CRPS metric is up to 60% for satellite measurements IMERG, 30% for radar data MRMS, and 10% for ground rain gauges. The company attributes this to training on high-quality sources, including NASA IMERG and its own precipitation reanalysis.
The model also takes into account weather station data, which helps more accurately predict local phenomena near coasts, mountains, and valleys. Additionally, WeatherNext 3 calculates wind speed at 100 meters height, cloud cover, and solar radiation — these parameters are important for planning energy production at wind and solar power plants.
Forecasts are already being used in Google Search, Gemini, and Google Maps. According to the company, for planning a day or more ahead, precipitation accuracy has increased to 50%. Developers and researchers can access the data through BigQuery, Earth Engine, and Google Cloud Storage.
The benefit for the regions of Latin America, Africa, and the Asia-Pacific region, where previously there was a lack of expensive local supercomputer models, is particularly noted. According to Google, WeatherNext 3 brings high-accuracy forecasts to billions of people in these areas. However, the company reminds that official weather warnings should be obtained from local meteorological services.
Primary source: blog.google ↗