Recently, the research team led by Professor Zhang Jiankai from the College of Atmospheric Sciences, Lanzhou University, has published its latest research paper entitled GMG: A Video Prediction Method Based on Global Focus and Motion Guided in IEEE Transactions on Circuits and Systems for Video Technology, a well known international journal at the intersection of computer science and video technology. The College of Atmospheric Sciences, Lanzhou University, serves as the primary completion unit. Master’s student Du Yuhao is the first author, and Professor Zhang Jiankai is the corresponding author. This research also received joint support from cooperating institutions including Central South University of Forestry and Technology, Ludwig Maximilians Universität München (LMU Munich), and the University of Manchester.
Figure 1. Core architecture of the GMG (Global focus and Motion Guided) model
With the rapid development of artificial intelligence driven meteorological forecasting, accurately characterizing the spatiotemporal evolution of atmospheric systems has become a key research direction at the intersection of meteorology and artificial intelligence. Different from ordinary video scenarios, meteorological processes feature large scale teleconnections, strongly nonlinear evolution, and non rigid morphological variations. Cloud clusters, precipitation and circulation systems often undergo generation, development, dissipation and deformation during their movement, posing challenges to existing spatiotemporal prediction models relying on local convolution or window based attention mechanisms.
To tackle this problem, the team achieved breakthroughs from two perspectives: global correlation modeling and motion pattern characterization. The Global Focus Module introduces critical background information of meteorological fields with low computational complexity and enhances the model’s perception of long range correlations. The Motion Guided Module improves prediction performance for the evolution of complex weather systems by modeling local growth, temporal decay and morphological change processes. This study reformulates meteorological forecasting as video prediction for sequential image frames, providing a more efficient and general modeling framework for AI models to understand and predict complex atmospheric processes.
In recent years, Professor Zhang Jiankai’s team has built continuous research expertise across interdisciplinary areas including AI driven meteorological forecasting, video prediction and atmospheric dynamic process modeling, yielding a series of outputs ranging from general method innovation to real world meteorological scenario applications. Represented by the GMG framework, the team addresses common bottlenecks in meteorological spatiotemporal prediction and proposes the Global Focus and Motion Guided video prediction architecture, improving model capability to depict large scale teleconnections and synoptic scale evolutionary processes.
Building upon this foundation, the team further carries out scenario oriented model design for meteorological forecasting at diverse time scales. For the prediction of stratospheric sudden warming events, the GMRNN model is developed targeting key dynamic processes such as polar vortex displacement and splitting, enabling skillful forecasts of typical stratospheric sudden warming events up to 20 days in advance. Relevant findings are published in the renowned geoscience journal Geophysical Research Letters. For nowcasting of precipitation, the GMSA model integrating global spatiotemporal attention and motion decoupling mechanisms is proposed to handle complex evolutions including rapid movement, intensification and dissipation of precipitation systems. It demonstrates outstanding generalization capability and prediction accuracy across multiple public precipitation datasets, with results published in Applied Soft Computing.
The above mentioned series of studies is supported by the Joint Fund for Meteorology of the National Natural Science Foundation of China (Grant No. U2442211).
Paper link: https://ieeexplore.ieee.org/document/11361248
