LSCNet: An Adaptive Cloud Detection Network via Local–Global Spatial Context

The ubiquitous presence of clouds in optical remote sensing images (RSIs) degrades image quality. Thin and fragmented clouds often exhibit low contrast and diverse morphology in complex scenes, which poses a significant challenge for accurate detection. To address the current challenges in thin cloud detection, this study proposes an Adaptive Cloud Detection Network in Remote Sensing Images via Local-Global Spatial Context (LSCNet). It aims to process thin cloud features within global and local spatial contexts, thereby improving the accuracy and robustness of thin cloud detection in complex environments. Specifically, the network simulates a dual perspective by constructing a Mamba-based Multiscale Fusion (MDF) block. This block utilizes learnable fusion weights to adaptively integrate differential and complementary information, thereby capturing thin cloud variations across both spatial and spectral dimensions in RSIs. Additionally, we propose a Local Gated Mamba (LGM) block for detailed feature enhancement. This module utilizes a spatial gating mechanism inspired by Long Short-Term Memory to capture key thin-cloud features and suppress residual background noise. By fully leveraging the spatial structure and morphology of thin clouds and building connections between thin cloud features across different spatial scales, the module achieves precise segmentation of cloud and ground features, thereby boosting the classification performance for thin and thick clouds in identical observational contexts. Extensive experiments conducted on the L8-Biome dataset and the WHUS2-CD+ dataset demonstrate that our method outperforms other existing cloud detection methods.