Global-Frequency Synergy: A Novel Paradigm for Radar Echo Extrapolation via Attention and Fourier Convolution

Accurate radar echo extrapolation is critical for short-term weather forecasting, yet existing deep learning methods often suffer from echo ambiguity, intensity decay, and insufficient global context utilization. To address these limitations, this paper proposes Global-Frequency Spatiotemporal Long Short-Term Memory (GFST-LSTM), a novel model that integrates a global attention mechanism and Fourier convolutional modules into the Spatiotemporal LSTM (ST-LSTM) architecture. The attention module dynamically weights multi-scale spatiotemporal features by enhancing channel and spatial correlations, while the Fourier convolution module captures global periodic patterns via frequency-domain transformations. Evaluated on the Moving Modified National Institute of Standards and Technology database (Moving MNIST) benchmark and Jiangsu Province radar data sets (2019–2021), GFST-LSTM achieves a 22.9% improvement in Critical Success Index and 13.1% in Heidke Skill Score over Predictive Recurrent Neural Network at the 40 dBZ threshold. Notably, it excels in preserving strong echo regions during 60–120 min predictions, reducing positional bias by 6.6% compared to the Motion Gated Recurrent Unit (MotionGRU). Ablation studies confirm the synergistic effect of both modules, with the full model outperforming variants that lack either component.