Jihad El Sana

High-resolution remote sensing imagery is important for accurate semantic segmentation, but it remains costly to acquire in many practical settings. Super-resolution (SR) offers a computational alternative for enhancing lower-resolution inputs, yet its value is still commonly assessed using visual reconstruction metrics that do not necessarily reflect downstream task utility. In this work, we evaluate SR from a functional perspective by measuring its impact on semantic segmentation performance. We introduce a controlled degradation-andreconstruction protocol in which original high-resolution images are downsampled, reconstructed using classical interpolation and learned SR methods, and then evaluated using fixed segmentation models without retraining. Experiments on four remote sensing datasets show that resolution degradation consistently harms segmentation performance, especially at stronger downsampling factors. Learned SR methods generally outperform classical interpolation, although the level of recovery remains dataset- and model-dependent. Across the primary datasets, SRGAN provides the most consistent functional preservation, while RealESRGAN is strongest in some specific settings. Overall, the results show that visual reconstruction quality alone is not a reliable indicator of downstream segmentation utility.