Multi-scale adaptive graph convolution-based thick cloud removal method for optical remote sensing images

Cloud occlusion challenges remote sensing image processing by degrading quality and analysis accuracy. Existing cloud removal methods capture local features but struggle with global dependencies and cloud morphology, limiting detail restoration and consistency. To address this, we propose a multi-scale adaptive graph convolution generative adversarial network (MAGC-GAN), integrating a multi-scale adaptive graph convolution network (MAGCN) and an adaptive patch discriminator (APD). MAGCN enhances spatial dependencies using adaptive graph convolution, effectively reconstructing cloud-covered regions by capturing global contextual relationships. A multi-scale feature fusion mechanism enables adaptation to varying cloud thicknesses. APD improves fine-detail recovery by evaluating multiple local patches individually, using an adaptive affine transformation matrix. It also incorporates texture-aware and global consistency losses to restore high-frequency details while maintaining coherence. Compared to existing methods, MAGC-GAN significantly enhances cloud-occluded region restoration, particularly in detail recovery and precise cloud edge reconstruction.