HyperMM: Satellite Image Sequence Prediction via Hypergraph-Enhanced Motion Matrix

Satellite image sequence prediction is a fundamental yet challenging task in weather forecasting. Existing approaches often combine neural network models (e.g., RNN, CNN, transformer, and GAN) to capture spatial features and temporal state transitions. While expressive, these models frequently overlook higher order motion correlations, leading to ambiguities in motion estimation and loss of appearance details. To address this, we propose HyperMM, a novel method for satellite image sequence prediction using a hypergraph-enhanced motion matrix. Unlike prior methods that compute pairwise feature similarities between pixels or patches within the same module, HyperMM constructs an appearance-independent motion matrix enhanced by hypergraph neural networks (HGNNs) to model high-order spatial correlations among pixel groups more effectively. Furthermore, we introduce interframe dynamic attention in the trunk network to improve temporal feature extraction, and intraframe static attention to reduce appearance information loss. Experiments on the Fengyun-4B (FY-4B) satellite dataset show that HyperMM achieves state-of-the-art performance by retaining fine-grained appearance detail while having the superior ability to accurately predict motion trends.