Dual-scale spatial attention network for restoring cloud-corrupted remote sensing imagery

Remote sensing imagery is prone to interference from atmospheric conditions during practical observation processes, which can cause occlusion or degradation of critical information within the images, thereby leading to errors in subsequent analysis and decision-making. Therefore, a dual-scale spatial attention mechanism-incorporated generative adversarial network (DSS-GAN) is proposed in this paper for cloud removal in remote sensing imagery. Firstly, spatial feature extraction in remote sensing imagery is structurally prioritized in the proposed framework, while enabling receptive field expansion and parameter redundancy suppression. Furthermore, strip convolution is strategically employed to capture high-aspect-ratio object features, effectively restoring elongated structural information such as mountains, rivers, and roads in the imagery. Finally, a residual Fourier branch(RFB) is introduced to integrate high- and low-frequency residuals from cloud-corrupted imagery, preserving edge texture details through spectral-spatial feature retention. Experimental results demonstrate that the proposed method effectively preserves edge-texture details while achieving an 81% reduction in parameter redundancy, with significant enhancements observed across both quantitative metrics and visual quality assessments.