A Multi-Channel Satellite Cloud Image Multi-Step Prediction Model Based on Motion Awareness and Multi-Scale Compensation

Accurate multi-step satellite cloud-image prediction remains challenging because cloud fields exhibit displacement, deformation, and local appearance changes over time. This study proposes M-SimVP, a multi-channel prediction framework built on SimVP. A Motion-Aware Temporal Prediction (MATP) module models temporal changes and learns alignment fields from historical features, while a Multi-Scale Motion Compensation (MSMC) module transfers the learned alignment information to multi-scale features and image space. A generation-refinement pathway further complements deformation-based compensation. Experiments on FY-4B/AGRI CH1-CH3 sequences use six input frames to predict four future frames at 15-, 30-, 45-, and 60 min lead times. M-SimVP achieves an MSE of 0.00421, SSIM of 0.7604, PSNR of 23.76 dB, and Edge-MSE of 0.0180, reducing MSE and Edge-MSE by 12.29% and 16.67%, respectively, compared with SimVP. The model also outperforms recurrent, optical-flow/advection, and satellite-specific baselines. Ablation, channel-wise, and lead-time analyses confirm the contributions of the proposed components and show that the performance advantage is maintained across the 1 h forecast horizon, while complex cloud evolution remains challenging at longer lead times.