Investigating the Potential of Super-Resolution for Road Segmentation in Sentinel-2 Images

Abstract

Road segmentation from Sentinel-2 imagery is challenging due to its coarse 10 m spatial resolution, yet its global coverage makes it valuable for applications like disaster relief and infrastructure monitoring. Traditional segmentation methods rely on high-resolution data, but recent approaches have explored super-resolution to enhance the spatial resolution of Sentinel-2 images. This study investigates the potential of super-resolution to perform road segmentation at 62.5 cm resolution from single-image Sentinel-2 RGB data, bridging the resolution domain gap. Both the super-resolution and the segmentation models are trained on high-resolution data, making the task more difficult. We demonstrate that these models can generalize to low-resolution data and deliver usable results for various applications, particularly in regions lacking up-to-date high-resolution imagery.