Restoration of Defective Cloud Images from Meteorological Satellites

Meteorological satellites are inevitably disturbed in information transmission, which leads to missing bars on the cloud images and thus makes it impossible for meteorologists to give accurate weather predictions. The UNet-like networks have a hierarchical feature map to obtain rich multi-scale contextual features. In addition, they can enhance the image reconstruction process using jump connections between the encoder and decoder. Therefore, based on deep learning theory, this paper proposes a method to restore defective satellite images using UNet and Swin-U'Net, which are both in shape of “U”. Moreover, the work selects the infrared satellite cloud image ir107 in the SEVIR dataset as the base dataset, from which the defective dataset is artificially generated. By comparing the restoration results of the two networks, the work concludes that the UNet network performs better than the Swin-UNet network in the image restoration function, which verifies the effectiveness of the network in the image restoration function.