<p>Cloud coverage in optical remote sensing images impacts surface information accuracy, making efficient cloud detection algorithms essential for remote sensing applications. This paper focuses on effectively detecting fragmented clouds and thin cloud boundaries in complex scenes using RGB satellite thumbnails that have lost a significant amount of spectral information. To address these challenges, we propose the Encoder-Decoder Network for Cloud Detection in Remote Sensing Thumbnails (CED). The CED comprises a Dual-Path Encoder (DPE) for feature extraction, a progressive Fusion Decoder (PFD) for cloud mask restoration, and Multi-Outputs for distributed supervision prediction. The DPE uses a CNN and a Transformer in parallel to extract hierarchical features at different scales. The CNN path extracts low-level features containing spatial structural information, while the Transformer path extracts mid-level features at various scales through Multi-Scale Dilated Attention (MSDA) and captures global high-level semantic features by establishing long-range dependencies through Multi-Head Self-Attention (MHSA).The PFD adopts a bottom-up structure and gradually fuses high-level semantic features from lower layers with multi-level features extracted by the encoder through Skip-Layer Fusion Blocks (SLFB), achieving progressive decoding and restoration of the cloud mask. To supervise the model more effectively, we employ Multi-Outputs for distributed supervision during training, enabling comprehensive and efficient cloud detection supervision. We also establish a Fusion Loss combining Binary Cross-Entropy (BCE)Loss and Dice Loss to simultaneously focus on class discrimination and similarity, which is beneficial for fine-grained cloud prediction. The CED was validated on the GF1-WHU and SPARCS datasets, achieving a Jaccard Index of 87.36% on the GF1-WHU dataset and 84.08% on the SPARCS dataset.</p>