S3-ESRGAN: Enhanced super-resolution generative adversarial network for remote sensing imagery spatial resolution improvement: An application using Sentinel-2 and UAV images

This study introduces a deep learning-based super-resolution approach to enhance the spatial resolution of Sentinel-2 imagery using uncrewed aerial vehicle (UAV) data. The proposed scale-adaptive, spatial-attentive, and spectral-preserving enhanced super-resolution generative adversarial network (S ³-ESRGAN) extends ESRGAN through three innovations: 1) a scale-adaptive mechanism designed for the 5× upscaling factor between images, 2) a spatial attention module emphasizing salient spatial and structural features, and 3) a spectral preservation component maintaining spectral and radiometric consistency. The architecture integrates dense blocks with spatial-attentive enhancement, multiscale feature extraction, and combined channel-spatial attention mechanisms. S ³-ESRGAN was trained on the Sen2UAV dataset containing over 428 000 Sentinel-2 and UAV patch pairs from 33 geographically diverse scenes. The composite loss function combines pixel-wise, perceptual, adversarial, spectral, and edge-preservation terms. Experimental results demonstrate that S ³-ESRGAN consistently outperforms state-of-the-art methods, producing sharper textures, enhanced edges, and improved spectral fidelity while preserving radiometric integrity. Ablation analysis further validates the contribution of the spectral preservation module to spectral stability. This research contributes to the remote sensing community by enabling the cost-free generation of high-resolution imagery from open-access satellite data.