PES-UNet: A Physics-Inspired Enhanced Hybrid Network for Cloud Classification From FY-4A Satellite Data

Cloud types are generated by various atmospheric thermodynamic and dynamic processes, reflecting the current atmospheric state and predicting future weather changes. Ac curate cloud classification is crucial for climate prediction. With the increasing demand for accurate cloud classification in remote sensing, CNNs have emerged in recent years as an essential computational tool. However, cloud shapes can change dramatically with environmental conditions, making it challenging to capture and recognize complex cloud boundaries. Additionally, CNNs often overlook the contextual features of clouds during their evolution, which results in the omission of fine-scale structures across various cloud types and compromises classification accuracy. To overcome these limitations, we introduce a physically constrained, interactive cloud classification framework—PES-UNet—that per forms pixel-wise semantic segmentation on satellite cloud imagery while simultaneously refining relevant physical parameters. First, the UNet network is employed to estimate complex atmospheric transmittance images by leveraging dark channel prior, which facilitates the preliminary delineation of cloudy versus cloud free regions. Subsequently, we design a feature fusion module that enables bidirectional interaction between CNN and Swin Transformer representations, ensuring the integration of detailed local features with comprehensive global semantics, facilitating fine-grained cloud type classification based on cloud region judgment. To evaluate the proposed architecture, we carry out cloud classification experiments using preprocessed FY4A L1 data covering most of China. The results demonstrate that, compared to other CNN, Transformer, or hybrid networks, PES UNet significantly enhances the accuracy of cloud classification in FY4A remote sensing imagery, achieving the highest match with Himawari-8 products. The network's ability to classify each pixel with semantic precision highlights its strength in considering both the fine-grained structure and the broader context of clouds during their evolution.