Thick Cloud Removal from Remote Sensing Images Using Double Shift Networks

Clouds greatly reduce the available ground information in optical remote sensing images. This paper proposed a double shift network to remove the thick clouds from multitemporal remote sensing images. The proposed networks are divided into two shift steps. In the first shift, the moment match and style transfer play the role of multi-temporal image normalization to obtain more reliable training images. In the second shift, in order to improve the network architecture's ability of capturing global semantics and local details, the shift connection layer and depthwise separable convolutions are introduced into U-Net. These two shift steps can not only improve the visual effect of cloud removal, but also further improve the quantitative evaluation. Experiments prove that the double shift network shows great advantages in cloud removal.