High-Definition from Above : Advancing Satellite Imagery Super-Resolution via Self-Supervision

This thesis contributes to the field of Earth Observation (EO) by advancing super-resolution techniques in satellite imagery, tackling both multi-image and single-image challenges. A key focus is overcoming the prevalent issue in remote sensing: the absence of high-resolution ground truth data. We address this through innovative self-supervised learning methods, which allow for model training without conventional high-resolution targets. In the realm of multi-image super-resolution, our Deep Shift-and-Add (DSA) framework, designed for SkySat sensors, exemplifies a novel self-supervised approach. It effectively combines shift-and-add fusion with self-supervised learning to handle the absence of high-resolution data, setting a new standard in the field. For single-image super-resolution, we leverage the unique properties of Sentinel-2 imagery to develop specialized self-supervised techniques. These methods enhance detail recovery, exceeding the sensor's inherent resolution. Our research not only addresses specific challenges in satellite imagery super-resolution but also hints at the broader applicability of our methods across various satellite platforms and data types, offering a promising direction for future exploration in satellite image processing and EO technologies.