ASENet: Thin Cloud Removal Network for Complex Scenes via Atmospheric Scattering Modeling and Feedback Enhancement

In optical remote sensing, thin clouds pose a significant challenge for cloud removal due to their high brightness and spectral similarity to bright man-made objects such as buildings. Existing thin cloud removal methods typically rely on single feature extraction or fixed physical model, which struggle to differentiate thin clouds from bright backgrounds in complex scenes, resulting in suboptimal image recovery. To address these issues, we propose ASENet, a novel network that integrates atmospheric scattering modeling with multi-level feedback enhancement mechanism to improve thin cloud removal for complex scenes. By learning the shape details of both thin clouds and ground features, ASENet dynamically adjusts weights in high-concentration cloud regions, ensuring clearer image recovery. Specifically, we design a feature fusion residual dehazing generator (FFRDG), which leverages deep residual blocks and high-resolution dehazing modules to capture environmental memory and enhance detail features, improving the model's adaptability and recovery accuracy in thin cloud regions. Additionally, to better preserve the edges and textures of buildings and other ground objects, we introduce a spatial detail enhanced discriminator (SDED) that incorporates the cascaded feedback-based feature mapping. This enables ASENet to better capture image details, maintain structural consistency, and effectively distinguish thin clouds from high-reflectance background objects. Extensive experiments on three benchmark datasets L8-ImgSet, RICE1, and WHUS2-CR demonstrate that our proposed ASENet outperforms state-of-the-art methods across both subjective and objective evaluation metrics, proving its effectiveness in thin cloud removal tasks under complex scenes.