Physics-Data-Coupled Network for Typhoon Satellite Cloud Image Sequence Prediction

Abstract

Accurate prediction of typhoon satellite cloud image sequences is essential for short-term typhoon monitoring and forecasting, including center localization, intensity estimation, and evolution analysis. Although deep learning methods have been widely used for typhoon satellite cloud image sequence prediction, they often fail to preserve key cloud structures and maintain motion-coherent evolution. To address these issues, we propose a physics-data-coupled network, termed PhyDC-Net, for typhoon satellite cloud image sequence prediction, in which an architectural design inspired by the transport–source structure of the atmospheric continuity equation is introduced to structure latent-state evolution. The proposed framework consists of two key modules. We design a physics-inspired gated recurrent unit (Phy-GRU) for the recurrent prediction layer. Inspired by the atmospheric continuity equation, this unit decouples the evolution process into a trend-dominated advection term and a transient-variation-driven residual correction term. This design introduces a physics-inspired inductive bias into recurrent prediction. To better capture the complex spatiotemporal evolution of typhoons, we further design a spatial-frequency spatiotemporal LSTM (SFT-LSTM) unit. SFT-LSTM replaces the conventional convolutional feature extractor in ST-LSTM with an integrated dual-branch feature extraction module whose outputs are used for subsequent gated updates. By combining hierarchical residual connections with the discrete cosine transform (DCT), SFT-LSTM jointly models spatial- and frequency-domain features, which improves the representation of fine cloud structures and rapidly varying patterns. Quantitative and qualitative experiments on the Digital Typhoon dataset show that PhyDC-Net outperforms representative baseline methods across multiple evaluation metrics, achieving an SSIM of 0.8138 and reducing mse by 17.2%. The results indicate that PhyDC-Net captures the large-scale motion of typhoon cloud systems while better preserving fine-grained structures, such as the typhoon eye, leading to more accurate and stable predictions.