Combined deep prior with low-rank tensor SVD for thick cloud removal in multitemporal images

Abstract

The thick cloud coverage phenomenon severely disturbs optical satellite observation missions (covering approximately 40–60% areas in the global scale). Therefore, the manner by which to eliminate thick cloud in remote sensing imagery is greatly significant and indispensable. In this study, we combine the deep spatio-temporal prior with low-rank tensor singular value decomposition (DP-LRTSVD) for thick cloud removal in multitemporal images. On the one hand, DP-LRTSVD utilizes the low-rank characteristic of multitemporal images via the third-order tensor SVD and completion. On the other hand, DP-LRTSVD employs the deep spatio-temporal feature expression ability by 3D convolutional neural network. The proposed framework can effectively eliminate thick cloud in multitemporal images through combining the model-driven and data-driven strategies. Moreover, DP-LRTSVD outperforms on thick cloud removal in the simulated and real multitemporal Sentinel-2/GF-1 experiments compared with model-driven or data-driven methods. In contrast with most methods that can only use a single reference image for thick cloud removal, the proposed method can simultaneously eliminate thick cloud in time-series images.