Super-resolution of Sentinel-2 imagery for remote sensing applications

This thesis investigates the application of deep learning-based super-resolution (SR) techniques to enhance the spatial resolution of Sentinel-2 satellite imagery, aiming to improve its utility across a range of remote sensing applications. With native spatial resolutions ranging from 10 to 60 meters, Sentinel-2 imagery often lacks the granularity required for detailed analysis in domains such as environmental monitoring, precision agriculture, urban planning, and disaster response. By leveraging recent advances in convolutional neural networks (CNNs) and generative adversarial networks (GANs), including architectures like SRCNN, FSRCNN, and ESRGAN, this work evaluates the effectiveness of various SR algorithms through both quantitative metrics (PSNR, SSIM, RMSE) and qualitative assessments across multiple test sites, including Sardinia (Italy), Lake Victoria (Kenya), and Milan (Italy). The research demonstrates that super-resolved imagery not only approximates very high-resolution (VHR) commercial data in structure and spectral characteristics but also supports improved ecological and urban analyses. A novel component of this work includes the application of SR-enhanced imagery to compute biodiversity indices in urban environments, highlighting the potential of enhanced EO data for computational ecological assessment. Ultimately, the thesis underscores the transformative potential of SR in making medium-resolution public imagery more actionable, bridging the gap between open-access and commercial data sources, and paving the way for scalable, cost-effective Earth observation solutions.