Evaluating the Reliability of Super-Resolution Satellite Imagery for Infrastructure Mapping: A Comparative Study of AI-Enhanced Sentinel-2 and High-Resolution Drone Imagery
Amri Rasyidi (a*), Yanuar Adji Nugroho (a)

a) ESRI Indonesia, Jalan Jend. Gatot Subroto Kav 18. Jakarta Selatan 12710, Indonesia
*amri.geodesy[at]gmail.com

Abstract

Artificial intelligence-driven super-resolution techniques generate visually convincing high-resolution satellite imagery from coarser inputs, yet their reliability for operational GeoAI applications remains scientifically unvalidated. While super-resolution methods can enhance 10.0 m Sentinel-2 L2A imagery to appear equivalent to 1.0 m resolution, the accuracy of AI-generated pixels for downstream semantic segmentation tasks lacks empirical assessment. This study quantifies performance boundaries between super-resolved Sentinel-2 using enhancement algorithms and native 1.0 m resolution drone imagery for infrastructure mapping applications, particularly road extraction. Three datasets covering identical geographic areas with temporally synchronized acquisition dates are compared using CNN-based and Transformer-based models architectures. Performance evaluation employs mean Intersection over Union (mIoU) and boundary accuracy metrics to assess both segmentation quality and edge preservation critical for infrastructure applications. Road extraction serves as a key test case, as typical road widths (&#8805-5.0 m) provide optimal conditions for evaluating super-resolution effectiveness against native 10.0 m Sentinel-2 limitations. Results demonstrate measurable performance difference in super-resolved imagery compared to native high-resolution data, with accuracy losses varying by model architecture. This research establishes evidence-based decision criteria for practitioners choosing between cost-effective super-resolution enhancement and native high-resolution acquisition, contributing to responsible deployment of AI-enhanced satellite imagery in operational remote sensing workflows.

Keywords: super-resolution, semantic segmentation, Sentinel-2, drone imagery, reliability assessment

Topic: Topic A: General Remote Sensing