On Training Deep Networks for Satellite Image Super-Resolution

The capabilities of super-resolution (SR) reconstruction (i.e., techniques for enhancing image spatial resolution) have been boosted recently by the use of deep convolutional neural networks. For SR, they are learned using huge training sets composed of original images, each of which is coupled with a low-resolution counterpart. In this paper, we explore how the SR performance depends on the procedure employed to obtain the training data. Up to date, this has not been given much attention—commonly, bicubic downsampling is used. Our extensive experimental study indicates that the training data characteristics have a large impact on the reconstruction accuracy, and the widely-adopted approach is not the most effective for dealing with satellite images. Overall, we argue that developing better training data preparation routines may be pivotal in making SR suitable for real-world applications.