Multiple-image super-resolution reconstruction using deep learning: A Sentinel-2 case study

Super-resolution reconstruction embraces a variety of techniques aimed at generating a high-resolution image from a low-resolution input. The goals of super-resolution range from hallucination that consists in producing a visually-attractive high-resolution image to reconstructing the real high-resolution information that is commonly required in Earth observation scenarios. The latter can be achieved by fusing a number of images with the same georeference, captured at a different time. Such multi-image super-resolution solutions underpinned with deep learning have been recently proposed for enhancing images acquired with the Proba-V sensor. This satellite captures images of 100 m and 300 m ground sampling distance, which makes it relatively easy to collect sufficient amounts of real-world data that can be used for training deep convolutional neural networks. As such data are unavailable for most of other satellites, alternative ways for training the networks must be adopted. In this paper, we demonstrate our solution for super-resolving multispectral Sentinel-2 images, and we present its most important components that may help implement multi-image super-resolution for other satellites. Our main focus is on the data that are used for training, as well as on the low-resolution images that are presented for reconstruction. We expect that the reported techniques will allow for increasing the capabilities of using Sentinel-2 images in a variety of practical Earth observation scenarios, but even more importantly, the presented methodology may be helpful in exploiting multiple-image SR for enhancing images captured with other satellites.