Relative performance of super-resolved Sentinel-2 and Copernicus VHR images in mapping built-up areas and building footprints using deep learning

Studies have demonstrated that impressive super-resolution results can be achieved on spaceborne optical images, such as Sentinel-2, using variants of the Generative Adversarial Networks (GAN) among others. However, the practical performance of these super-resolved images in various applications, co...

Full description

Saved in:
Bibliographic Details
Main Author: Misganu Debella-Gilo
Format: Article
Language:English
Published: Taylor & Francis Group 2025-12-01
Series:European Journal of Remote Sensing
Subjects:
Online Access:https://www.tandfonline.com/doi/10.1080/22797254.2025.2517381
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Studies have demonstrated that impressive super-resolution results can be achieved on spaceborne optical images, such as Sentinel-2, using variants of the Generative Adversarial Networks (GAN) among others. However, the practical performance of these super-resolved images in various applications, compared to their original high-resolution and low-resolution counterparts, is less researched. This study explores enhancing the spatial resolution of spaceborne remote sensing images from the Sentinel-2 satellite, using deep learning-based super-resolution methods. It then evaluates the performance of the super-resolved images against Very High-Resolution (VHR) optical satellite images and original Sentinel-2 images (S2) for mapping built-up areas and building footprints. A GAN-based single-image super-resolution model was trained from scratch on S2 tiles to enhance the resolution by a factor of five, creating three datasets: original S2 (10 m resolution), their super-resolved version (S2-SR), and VHR images (2 m resolution). Using a U-Net semantic segmentation model, the study mapped built-up areas and building footprints, employing both raster-based and vector-based evaluation metrics. Results show that S2-SR images perform comparably to VHR images in mapping built-up areas and slightly lower in mapping building footprints but significantly outperform original S2 images in both tasks.
ISSN:2279-7254