Precipitation nowcasting represents a critically important task in meteorological domain, fundamentally characterized by the spatiotemporal extrapolation of radar echo images. The prevailing methods predominantly rely on integrated frameworks combining Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs). Nevertheless, owing to the inherent limitations of convolutional operations in modeling long-range spatial dependencies, these methods exhibit deficiencies in holistically capturing the dynamical evolution of precipitation systems and underestimate the intensity and spatial extent of severe precipitation scenarios. To address these constraints, this paper proposes the WinG-LSTM model, which achieves synergistic integration of the Swin Transformer model and the PredRNN model. Notably, the Swin Transformer incorporates a gating mechanism within its Multi-Layer Perceptron (MLP) blocks to augment feature representation capabilities. Rigorous evaluation conducted on the CIKM2017 dataset and Shanghai Radar dataset demonstrate the proposed model’s superior performance relative to all comparative models. Quantitative and qualitative analyses confirm that WinG-LSTM generates predictions exhibiting significantly higher structural similarity to ground-truth radar observations, thereby delivering enhanced accuracy in precipitation nowcasting.
Yang et al. (Sat,) studied this question.