"Identification of South China Sea Fog from FY-4A AGRI Imaging Using Multi-Task Learning with Physical Constraints and Cross-Scale Consistency"

"Sea fog identification remains challenging due to the limited physical interpretability of deep learning models and the unclear interaction mechanisms in multi-task learning frameworks. To address these issues, this study proposes a physics-informed multi-task learning method (PIMTL) for sea fog identification based on multi-source data, with FY-4A AGRI satellite imagery as the primary observational input. The high-temporal-resolution geostationary observations provided by FY-4A AGRI enable continuous monitoring of sea fog evolution, offering rich spectral information for feature construction. An encoder\u2013decoder architecture is constructed based on ResNet-50 and U-Net. In addition to the primary segmentation task, pixel-level physical auxiliary tasks, including liquid water path (LWP) and fog top height (FTH), are introduced, together with semantic tasks such as fog intensity estimation and dense fog classification, enabling unified multi-task modeling. Systematic ablation experiments reveal that pixel-level physical tasks enhance segmentation performance by providing consistent spatial constraints, whereas sample-level semantic tasks may introduce negative transfer due to mismatched supervision granularity. To mitigate optimization conflicts among tasks, a stage-wise training strategy is further proposed, in which task weights are dynamically adjusted across different training stages, allowing a smooth transition from multi-task collaborative learning to task-specific refinement. Experimental results demonstrate that the proposed method achieves an accuracy of 0.8953 while reducing the false alarm rate by approximately 23.6% compared with conventional models, and exhibits superior stability and physical consistency. Cross-scale validation using NOAA station observations further confirms that the proposed approach maintains robust generalization capability under complex air\u2013sea interaction conditions, highlighting its effectiveness for operational sea fog monitoring based on FY-4A AGRI observations."