Super Resolution on Sentinel-2 RGB images using Deep Learning algorithms

Super resolution (SR) techniques are being widely studied and applied on optical images, providing interesting results in the context of image processing. When applied to earth observation (EO) applications, SR is able to provide images with an increased spatial resolution, enabling a quality enhancement of all the postprocessing techniques which are employed over the super-resolved images. This work represents a study performed on Sentinel-2 RGB images focused on enhancing their spatial resolution using a Deep Learning (DL) based approach. DL algorithms have shown great interest in the satellite remote sensing domain both for the high variety of applications that this kind of data has and for the continuous evolution of the satellite sensors. To solve the Super Resolution problem, Deep learning state of the art algorithms have been employed. Moreover, the core technology is represented by Generative adversarial networks that, according to literature, have proven to outperform many SR benchmarks. The first part of the work is devoted to the analysis of the SR task and its advantages on remote sensing applications. Moreover, a background description related to the data involved in the work and their management is provided. To fully understand the proposed SR methods, the exploited Deep learning technologies and some essential task specific tools like metrics or classical up sampling methods are deepened in the second section. The third and fourth chapters are the main core of this work, where two SR algorithms are presented. The first one takes the name of Enhanced Super-Resolution Generative Adversarial Networks (ESRGAN) and it is the starting point of the second algorithm, known as DKNSRGAN. The DKNSRGAN aims to solve some drawbacks experienced in the first approach that is mainly related to the lack of generality when the model performs on real Sentinel-2 RGB images. This problem stems from the method through which input-label training data couples are generated. The algorithm proposes a more structured method to perform this operation employing also a KernelGAN model that has the role of producing natural downscaled images. Moreover, the algorithms are tested on a specifically built dataset and all the results are reported and analyzed in depth. One possible advantage of developing a SR model capable of enhancing the resolution of EO images is to use it as a preprocessing step before object detection tasks. Nowadays, for real-time monitoring scenarios, this task can also be performed on-board the satellite platform, thanks to the computational power of the computing devices available on the market. As a result, one section is also devoted to model optimization, where the possibility of running the SR model developed in this work on an embedded device is assessed. In the last chapter a comparison of the overall results that have been obtained in this work is performed. Furthermore, some insights regarding future improvements and further analysis are presented.