DIF-UNet: deep convective cloud recognition from FY-4A multi-channel observations

Deep convective clouds are frequently associated with extreme weather events such as torrential rainfall and thunderstorms, making their accurate identification critical for improving short-term weather forecasting and enabling effective disaster early warning. Most existing deep learning approaches for cloud classification rely on conventional convolutional neural networks (CNNs), which suffer from limited receptive fields and struggle to capture long-range contextual dependencies essential for robust cloud segmentation. To overcome this limitation, we propose DIF-UNet, a novel U-Net architecture that combines a Swin Transformer and a CNN-based residual network as dual encoders. The model introduces two key components: (1) a Semantic Information Fusion Module (SIFM) that adaptively integrates heterogeneous semantic features from both encoders using learnable parameters and (2) a Cross-Layer Feature Interaction Module (CLFI) that enhances hierarchical feature synergy to preserve fine-grained spatial details while reducing information loss. Evaluated on FY-4A multi-channel satellite observations, DIF-UNet achieves state-of-the-art performance with an intersection-over-union (IoU) score of 77.78% and an F1 score of 87.39%, outperforming baseline models by 3.84% and 2.49%, respectively. These results demonstrate that explicitly modelling both local and global contextual information significantly improves deep convective cloud recognition accuracy.