Google Unveils 'WeatherNext 3' Based on Real-Time Satellite Data… 5km High-Resolution and Hourly Weather Forecasts

Google DeepMind and Google Research have released WeatherNext 3. This model directly learns from real-time satellite data instead of relying on traditional physics simulations, providing hourly and regional forecasts that are five times sharper than previous versions.

Architecture and Resolution Improvements

WeatherNext 3 utilizes a single Functional Generative Network (FGN) architecture, taking geostationary satellite mosaics and traditional historical reanalysis data as inputs. This enables predictions with 5km resolution for 2m surface temperature, 10km for other surface variables, and 25km for atmospheric variables such as wind speed. It eliminates the excessive pixelated smoothing seen in previous models and clearly distinguishes complex regional terrain.

Improved Precipitation Forecast Accuracy

By learning from NASA's IMERG and proprietary satellite radar-based reanalysis data, precipitation forecast accuracy has been significantly improved. Based on the medium-range global forecast evaluation (CRPS), it shows up to 60% improvement over IMERG and 30% over MRMS, with a 10% accuracy improvement over rain gauge measurements at initial lead times.

Usage and Accessibility

The model is utilized in agriculture, renewable energy, and daily planning, and is directly accessible via Google Search, Maps, and Gemini. Additionally, data integration is possible through Google Cloud, with queries available in BigQuery and Earth Engine. It provides high-quality data to regions such as Latin America and Africa, which previously lacked high-resolution forecasts due to supercomputing costs.