Shadow removal in high-resolution satellite images using conditional generative adversarial networks

In satellite image processing, obscure zones that were affected by shadows are normally discarded from further processing. Nevertheless, for specific applications, such as surveillance, it is desirable to remove shadows despite the fact that reconstructed zones do not necessarily have real reflectance values. In that sense, we propose a shadow removal method in high-resolution satellite images using conditional Generative Adversarial Networks (cGANs). The generator network is trained to produce shadow-free RGB images with condition on a satellite image patch altered with artificial shadows and concatenated with its respective binary shadow mask, while the discriminator is adversely trained to discern if a given shadow-free image comes from the generator or if it is an original RGB image without artificial alteration. The method is tested in the proposed dataset achieving an error ratio comparable with the state of the art. Finally, we confirm the feasibility of the proposed network using real shadowed images.