Proba-V Super-Resolution in combination with Sentinel-2

In the field of Deep Learning, and more specifically Computer Vision, one of the tasks that covers quite a lot of interest is that of Super-Resolution: a set of techniques used to improve the resolution of digital images. Super-Resolution techniques have found room for application in many fields, among which one of the most interesting is that of remote sensing and earth observation; this is amply evidenced by the very numerous challenges on the subject organized by space entities such as ESA and NASA. Being able to improve the spatial resolution, i.e. the physical measurement (meters) that represents the size of a pixel, of a satellite image can be particularly useful for a number of reasons including being able to make object classification and detection tasks easier to solve, or again, to monitor at a greater level of detail the earth's surface. High spatial resolution images, however, are produced by remote sensing satellites less frequently than low spatial resolution ones, which is why having models available that allow one to increase resolution from one or more low resolution images turns out to be a critically important task. PIUnet is a model that performs MultiTemporal Image Super Resolution, created by a research group at Politecnico di Torino as a result of a challenge convened by ESA, which increases the spatial resolution of images from the Proba-V satellite from 300 to 100 meters. Starting from the existing Super-Resolution architecture PIUnet, the purpose of this thesis is to evaluate how the Super-Resolution task can be performed using images from two different missions, Proba-V and Sentinel-2. Specifically, we want to make sure that the low-resolution model training images remain those from Proba-V while the ones used as Ground-Truth (i.e., high resolution) are instead selected from Sentinel-2. To do this, the work presented was divided into several phases: a first phase of creation of the dataset comprising low-resolution images from Proba-V and high-resolution images from Sentinel-2, a second phase consisting in the training of the model on the new dataset, a third phase of evaluating the results obtained from the standard version of PIUnet, and finally a fourth phase of modifying PIUnet to make it suitable for working with images from two different satellites. Each of these stages is described in details within this thesis. Furthermore, at the end of this work, both the results obtained using the standard version of PIUnet and the results obtained from the modified version of PIUnet are also presented, analyzed, and compared in depth, showing how the changes made on PIUnet resulted in better quality, radiometrically-consistent, super-resolved images.