Evaluating sentinel-2 super-resolution algorithms for automated building delineation

Abstract

Super-Resolution is the result of enhancing the spatial resolution of images while introducing high-resolution details and has been proposed as a way to make satellite data such as Sentinel-2 more useful for real-world applications. It promises higher-resolution imagery at a low cost, as existing image sources can be relied on. While many Super-Resolution models with varying architectures and training datasets have been developed, independent frameworks to evaluate their effectiveness in real-world tasks remain limited. This thesis examines the performance of Super-Resolution for Sentinel-2 imagery on the downstream task of building delineation using a novel high-quality dataset covering Austria, created specifically for the proposed task. The methodology consists of three main steps: first, constructing a spatially and temporally aligned dataset with orthophotos as high-resolution references, Sentinel-2 images as lower-resolution inputs, and cadastral masks as ground-truth labels; second, applying Super-Resolution models to super-resolve Sentinel-2 images from a spatial resolution of 10 meter to 2.5 meter. In parallel, the Sentinel-2 images are upsampled to 2.5 meter using interpolation methods to provide a deterministic baseline for evaluating the Super-Resolution results; and third, training UNet models for building delineation on both super-resolved outputs and interpolated Sentinel-2 images. Three main findings emerge: first, orthophoto-based building delineation achieves the best results; second, models trained on interpolated images outperform those using Super-Resolution outputs; third, the differences between Super-Resolution models demonstrate that their choice of training data and architecture has a considerable influence on performance. These results suggest that, for building delineation, Super-Resolution currently offers no advantage over interpolation methods, and further development is needed to justify its real-world application. Furthermore, using original high-resolution image sources allows the most accurate result, if such data is available. More broadly, the outcomes highlight the importance of structured evaluation frameworks for benchmarking Super-Resolution on downstream tasks. Extending such evaluations to include diverse real-world applications will be essential for advancing Super-Resolution towards robust and task-oriented models capable of delivering reliable super-resolved satellite imagery.