GLGF-CR: A Gated Local-Global Fusion approach for cloud removal in real-world remote sensing

Optical satellite imagery is a critical data source for Earth observation in remote sensing. However, cloud cover often degrades image quality, hindering its application and analysis. Therefore, effective cloud removal from optical satellite images has become a prominent research direction. In real-world scenarios, thick clouds act as pure noise, completely obscuring underlying information, while thin clouds provide partially beneficial information that can be leveraged for reconstruction. Traditional cloud removal methods often fail to distinguish between these two types of noise, leading to suboptimal performance. To address this limitation, we propose a novel cloud removal model, GLGF-CR, which incorporates a Gated Local-Global Fusion module. This module is designed to effectively separate and process the distinct characteristics of thick and thin clouds. For thick clouds, which contain no recoverable information, the model focuses on robust reconstruction using complementary data sources. For thin clouds, the model extracts and utilizes the beneficial information embedded in the partially obscured regions, enabling more accurate and detailed reconstruction. Additionally, a Dual Cross-Attention mechanism is introduced to establish robust mappings between SAR and optical modalities, further improving fusion accuracy. To handle domain shifts between source and target domains, we incorporate a domain adaptation module, which enhances the model's ability to generalize across diverse real-world scenarios. The proposed algorithm not only outperforms existing methods on the large-scale real-world dataset SEN12MS-CR but also demonstrates strong cross-domain transferability on the Henan flood dataset. By explicitly addressing the dual nature of cloud noise–pure noise in thick clouds and partially beneficial information in thin clouds–this work advances the field of beneficial noise learning, demonstrating how noise can be systematically analyzed and utilized to improve model performance in complex scenarios.