Deep Image- To-Image Transfer Applied to Resolution Enhancement of Sentinel-2 Images

Single Image Super-Resolution (SISR) is looking at restoring the missing high-resolution information from a single low-resolution image in order to increase the apparent spatial resolution by a factor of two or more. In recent years, convolution neural networks have been applied with great success to the problem of improving spatial resolution from a single image. With the advent of low-resolution (10 m) optical sensors such as Sentinel-2, it is interesting to explore the possibility of improving image resolution with Deep Learning (DL) techniques. The purpose of this article is to investigate the potential performances of recent DL super-resolution techniques. The techniques explored here include not only techniques for enhancing high-frequency content but also so-called image-to-image translation techniques based on Generative Adversarial Neural Networks (GAN). From our preliminary results, we show that GANs have the ability to restore complex textural information.

1. Introduction

Nowadays, the Sentinel-2 constellation produces multispectral optical data at ten meters resolution. Common ways to improve spatial resolution in remote sensing includes the use of interpolation filters (e.g. bicubic) or pan-sharpening techniques when a higher resolution pan-chromatic is available.