Abstract:

Since remote sensing images will inevitably be affected by the climate in the acquisition process,obtained images may contain cloud information,which affects the subsequent use of images to a large extent.Image cloud removal methods based on deep learning can remove clouds well,but they have problems such as long training time,insufficient cloud removal effect and color distortion.To solve these problems,a cloud removal method based on the end-to-end generative adversarial network (GAN) is proposed to recover clear images from remote sensing images containing clouds.First,the U-Net is used as the main structure of the generator,and a continuous memory residual module is added between the encoder module and the decoder module to mine the depth characteristics of the input information.Then,a convolutional neural network is adopted as the discriminator to distinguish authenticity.Finally,the loss function,by combining the adversarial function with the absolute loss function,is designed to measure the advantages and disadvantages of the model by calculating the gap between the output of the network model and the real data.Experimental results show that the proposed method is superior to existing cloud removal methods in both quantitative indexes (peak signal to noise ratio and structural similarity) and running time.Under the same number of parameters,the proposed method has the lowest calculation amount (GFLOPs) and a lower algorithm complexity.Besides,remote sensing images obtained by the proposed method can lead to richer detailed information,almost no color distortion,and a better subjective visual effect.

Key words: remote sensing image, image cloud removal, generative adversarial network, continuous memory residual