Monitoring road infrastructure is essential for accessibility and economic development, yet data scarcity remains a major constraint in many developing countries. Satellite-based deep learning approaches provide a scalable alternative, but the coarse resolution of freely available imagery may limit performance. This study evaluates whether generative adversarial network-based super-resolution can improve road roughness classification from Sentinel-2 imagery. Using a national dataset of 124,462 road segments from the Philippines, we compare classification results from native medium-resolution inputs with high-resolution images generated using real-ESRGAN. Despite clear visual improvements and higher perceptual quality scores (PSNR increasing from 15.3 to 16.7 dB), super-resolution did not improve predictive performance. Binary classification accuracy differed by less than 0.5% points, while four-class accuracy declined by 2 to 4% points when using super-resolved imagery. Most errors occurred between adjacent roughness categories, and super-resolution did not reduce this confusion. In contrast, combining native imagery with environmental covariates, including temperature, precipitation, slope, and population density, substantially improved results, achieving 85% binary and 71% four-class accuracy. These findings indicate that preserving original spectral information and incorporating contextual variables is more effective than generative image enhancement for satellite-based road quality assessment in data-sparse settings.
Thegeya et al. (Thu,) studied this question.