Single-image cloud removal in optical remote sensing using hybrid CNN-ViT architectures

Abstract Cloud cover remains a major challenge in optical remote sensing, as it obscures critical surface features and degrades the reliability of environmental analysis. Many existing cloud-removal methods rely on multi-temporal or multi-sensor observations. However, such auxiliary data are often unavailable in time-sensitive operations or when only single-date satellite imagery is accessible. To overcome this limitation, this work proposes a lightweight cloud-removal framework based only on spatial cues to reconstruct clear land surfaces using only one cloudy input image. The proposed method integrates Convolutional Neural Networks (CNNs), specifically the Super-Resolution Convolutional Neural Network (SRCNN) and UNet with Vision Transformer (ViT) modules. A CNN component extracts fine local spatial details from the input image, whereas a ViT module captures global contextual information. Despite its computational complexity, the framework effectively models both local and global features. Mask-guided compositing ensures radiometric safety by preserving non-cloud regions, while region-aware evaluation measures reconstruction quality across cloud interiors, boundaries and clear zones. Compared to baseline CNN models, the ViT-enhanced architectures achieve superior structural fidelity, seam continuity and perceptual realism on the RICE-I (thin) and RICE-II (thick) datasets, which consist of paired cloudy and cloud-free optical images representing thin and thick cloud conditions. On thin clouds, UNet-ViT achieved 35.75 dB PSNR and 0.969 SSIM, while on thick clouds it attains 32.14 dB PSNR and 0.923 SSIM within cloud-masked regions, while preserving clear areas with negligible radiometric deviation (Δ ≈ 0).