UNROLLED LOW-RANK TENSOR COMPLETION FOR SAR-GUIDED THICK CLOUD REMOVAL IN HYPERSPECTRAL SATELLITE IMAGES

We propose a deep unrolling-based low-rank tensor completion algorithm for cloud removal in hyperspectral satellite images. First, by exploiting the intrinsic low-rank structure of hyperspectral images (HSIs) and complementary structural information from synthetic aperture radar (SAR) data, we formulate the cloud removal task as a joint optimization problem. This formulation integrates a low-rank tensor model, an SAR-guided prior for detail injection, learnable implicit regularizers, and an attention mechanism to emphasize salient spatial-spectral features. Then, we solve the optimization problem via a multistage deep unrolled network, where each stage updates the optimization variables using closed-form solutions and learned regularizers. Experimental results demonstrate that the proposed algorithm significantly outperforms existing algorithms, achieving superior reconstruction quality for cloudy HSIs.