CoBE-GAN: Correction of Banding Error Using Generative Adversarial Network

Abstract Remote sensing images are often used for scientific or military research, commercial activities and urban planning, etc. However, these data are susceptible to radiometric, geometric errors, and other distortions caused by atmospheric disturbances, affecting the consistency and the comparability of remote sensing images. Generative Adversarial Networks (GAN) is effective for complex remote sensing image error correction. Therefore, this paper proposed a CoBE-GAN (Correction of Banding Error - GAN) method to correct remote sensing image with banding error. In generator, we combine the self-attention (SA) structure with U-Net to enable the generator to integrate global information in the generation process and the model to better capture correlations between different locations in the image. In discriminator, we proposed a SA-Patch module, which integrates self-attention with PatchGAN, is designed to better focus on global information and improve the adversarial stability between the generator and the discriminator and make the training easier to converge. A total of 5 deep-learning networks were compared using the Landsat 7 ETM+ remote sensing images with banding errors in 2003. Our model achieved a mean Structural Similarity (SSIM) of 0.958 with a Peak signal-to-noise ratio (PSNR) of 41.624 dB and Mean Squared Error (MSE) 4.535, which is the best of the comparison models. Furthermore in visual comparison, the proposed model eliminating errors better in generated remote sensing images. The result reflected the superiority of our method in banding errors correction of remote sensing images.