CFPNet: Coarse-to-Fine Progressive Network for Cloud Detection in Remote Sensing Images

Accurate cloud detection in remote sensing images constitutes a critical preprocessing requirement for ensuring the validity of subsequent analytical applications. Existing cloud detection methods effectively identify primary cloud regions but struggle with precise boundary delineation and distinguishing spectrally similar surfaces, particularly between clouds and terrain with comparable reflectance properties. To address these challenges, we propose the coarse-to-fine progressive network (CFPNet), a progressive detection framework integrating across channel and spatial dimensions. The framework incorporates two innovative components: a channel spatial attention residual fusion (CSARF) module and a multimask adaptive attention (MMAA) module. The CSARF module achieves the network’s initial focus on clouds, while the MMAA module enables fine feature extraction of clouds. Specifically, The CSARF module uses channel attention module (CAM) and spatial attention module (SAM) to suppress background noise while enhancing discriminative feature representation. MMAA uses multimask self-attention (MMSA) to compute channel self-attention and capture long-range dependencies, while using a multimask strategy to filter important channels. Deformable contextual feedforward network (DCFN) then adaptively extracts cloud boundary features through deformable convolution, minimizing noncloud pixel interference. Hence, the network enables coarse-to-fine feature extraction of clouds across both channel and spatial dimensions. Experimental results on the GF1-WFV, AIR-CD, and Sentinel-2 datasets demonstrate that our method achieves superior performance in boundary detail accuracy and interclass feature classification compared with other methods.