Fully-connected tensor network decomposition and group sparsity for multitemporal images cloud removal

Existing nonblinded cloud removal approaches in multitemporal images depend on the accuracy of the mask used. However, masks are typically detected via manual labeling or cloud detection methods, which do not guarantee accuracy and thus may affect cloud removal. In this paper, we present a cloud/shadow removal method that does not require masks: fully-connected tensor network decomposition and group sparsity (FCTNGS). To capture multitemporal information, we use fully-connected tensor network decomposition to explore the global correlation of multitemporal images and weighted group sparsity to describe cloud sparsity. To develop the proposed model, we propose an efficient algorithm that is based on the proximal alternating minimization (PAM) method. Experiments with both simulated and real datasets demonstrate the effectiveness and robustness of the proposed cloud/shadow removal method.