Improving Satellite Image Fusion via Generative Adversarial Training

The optical images acquired from satellite platforms are commonly multiresolution images, and converting multiresolution satellite images into full higher-resolution (HR) images has been a critical technique for improving the image quality. In this study, we introduced the generative adversarial network (GAN) and proposed a new fusion GAN (FusGAN) approach for solving the remote sensing image fusion problem. Specifically, we developed a new adversarial training strategy: 1) downscaled multiresolution images are adopted for generative network (G-Net) training, and 2) the discriminative network (D-Net) is used to adversarially train the G-Net by discriminating whether the original multiresolution images have been fused well enough. To further improve the capability of the network, we structured our G-Net with residual dense blocks by combining state-of-the-art residual and dense connection ideas. Our proposed FusGAN approach is evaluated both visually and quantitatively on Sentinel-2 and Landsat Operational Land Imager (OLI) multiresolution images. As demonstrated by the results, the proposed FusGAN approach outperforms the selected benchmark methods and both perfectly preserves spectral information and reconstructs spatial information in image fusion. Considering the common resolution disparities among intra- and intersatellite images, the proposed FusGAN approach can contribute to the quality improvement of satellite images and thus improve remote sensing applications.