Snow or rain? hybrid AI deciphers surface precipitation phase from satellite observations

Area of Science:

• Meteorology
• Artificial Intelligence
• Remote Sensing

Background:

• Surface precipitation phase transitions cause severe weather events like snowstorms and avalanches.
• Accurate, real-time surface observation data is crucial but currently limited globally.

Purpose of the Study:

• To develop a novel AI framework for real-time quantification of surface precipitation phase using satellite observations.
• To overcome the latency issues associated with current operational precipitation phase monitoring systems.

Main Methods:

• Developed the Real-time Precipitation Phase-Intensity Collaborative Retrieval Network (RePPIC-Net), a hybrid AI framework.
• Integrated real-time 3D atmospheric physics fields from the AI-driven FuXi model with geostationary satellite data.
• Employed a hierarchical architecture for efficient data processing and analysis.

Main Results:

• RePPIC-Net enables real-time monitoring of surface precipitation phase, significantly reducing latency compared to existing systems.
• Validation against ground stations in China showed improved Critical Success Index scores for snowfall and rainfall detection.
• Achieved CSI scores of 0.1574 (snowfall) and 0.3147 (rainfall) for light precipitation (0.1-5 mm/h), surpassing operational products.

Conclusions:

• RePPIC-Net offers a breakthrough in real-time precipitation phase monitoring from space.
• The system's capability supports the development of satellite-based nowcasting for precipitation phase transitions.
• Provides a replicable model for AI-driven real-time weather monitoring, addressing critical gaps in winter weather disaster warnings.