A researcher from the National Satellite Meteorological Centre in China has developed a model that has reportedly extended the effective lead time of nowcasting for severe convective weather to four hours.
Wang Jingson, working with researchers from Harbin Institute of Technology (Shenzhen), Hong Kong University of Science and Technology and Guangzhou Institute of Tropical and Marine Meteorology at the China Meteorological Administration (CMA), engineered the deep diffusion model based on data gathered by the Fengyun (FY) meteorological satellites.
The CMA says it marks an important breakthrough in China’s intelligent forecasting technology based on independent satellite data and will provide stronger technical support for disaster prevention and mitigation.
Severe convective weather is characterized by sudden occurrence, rapid evolution, and devastating power. The key challenge of its nowcasting lies in capturing the rapid non-linear evolution of meso- and micro-scale systems. Accurately forecasting severe weather of this nature is a common challenge faced by those in the international meteorological community.
The research team found that severe convection mostly occurs in convective cloud systems or isolated convective cloud clusters. Equipped with high spatiotemporal resolution infrared detection capabilities, the FY-4 meteorological satellites provide continuous, large-scale observations of cloud top brightness temperature, enabling full tracking of the entire lifecycle of cloud clusters. By capturing physical signals from cloud tops, the satellites can detect early signs of convection initiation at an earlier stage, thus gaining valuable lead time for forecasting.
Leveraging the monitoring capabilities of the FY-4 satellite series, the research team obtained data which was applied to extract and predict the complex and random movements of convective cloud clusters from massive volumes of satellite data.
The deep diffusion model for satellite data (DDMS)
To improve forecasting accuracy, the team introduced the diffusion model, which it says demonstrated reliable performance in the field of image generation in recent years, and proposed a deep diffusion model for satellite data (DDMS).
This model simulates the random movement trends exhibited during the evolution of convective clouds as a physical diffusion process. It harnesses a two-hour sequence of infrared brightness temperature data from the FY-4A satellite to predict the spatio-temporal evolution of convective clouds over the next four hours.
On this basis, combined with a deep semantic segmentation model, the model performs automatic identification and spatial localization of convective clouds from the predicted satellite data sequences, accurately depicting the formation and development processes of convective clouds.
Based on data from the FY-4 satellites, the DDMS has reportedly achieved high-resolution convective forecasting for an area of approximately 20,000,000km² covering China and its surrounding regions, generating predictions every 15 minutes for the next four hours. It is also said to demonstrate stable forecasting performance across different spatial scales (4,000-48,000m) and seasons, while maintaining high reliability not only in short-term forecasting but also in extended lead-time forecasting ranging from two hours to four hours.
This research explores the integration of FY-4 satellite observations with a deep diffusion model, proposing a new technical framework for severe convective nowcasting. It also identifies avenues for future development and demonstrates the model’s potential to improve forecast timeliness, accuracy and reliability, thereby strengthening technical support for disaster prevention and mitigation of severe convective events such as heavy rainfall, thunderstorms and short-term gales.
In the future, the research team aims to further optimize the computational efficiency of the model, and plans to explore the deeper integration of meteorological physical laws into the AI model to enhance its interpretability and robustness.
In related news, China completes coordinated mountain snowfall observation experiment in North China
