Thick cloud and cloud shadow removal in multitemporal imagery using progressively spatio-temporal patch group deep learning

Thick cloud and its shadow severely reduce the data usability of optical satellite remote sensing data. Although many approaches have been presented for cloud and cloud shadow removal, most of these approaches are still inadequate in terms of dealing with the following three issues: (1) thick cloud cover with large-scale areas, (2) all the temporal images included cloud or shadow, and (3) deficient utilization of only single temporal images. A novel spatio-temporal patch group deep learning framework for gap-filling through multiple temporal cloudy images is proposed to overcome these issues. The global-local loss function is presented to optimize the training model through cloud-covered and free regions, considering both the global consistency and local particularity. In addition, weighted aggregation and progressive iteration are utilized for reconstructing the holistic results. A series of simulated and real experiments are then performed to validate the effectiveness of the proposed method. Especially on Sentinel-2 MSI and Landsat-8 OLI with single/multitemporal images, under small/large scale regions, respectively.