S ³ -ESRGAN: Enhanced Super-Resolution Generative Adversarial Network for Remote Sensing Imagery Spatial Resolution Improvement—An Application Using Sentinel-2 and UAV Images

This research proposed an innovative deep learning based super-resolution approach to enhance the spatial resolution of medium-resolution satellite (specifically Sentinel-2) imagery using unmanned aerial vehicle (UAV) images. The method named Scale-Adaptive, Spatial-Attentive, and Spectral Preserving Enhanced Super-resolution Generative Adversarial Network (S³-ESRGAN) is a novel architecture that enhances the ESRGAN framework with three innovations. The innovations include: (1) a scale-adaptive mechanism specifically designed for the 5× upscaling factor between Sentinel-2 and UAV imagery, (2) a spatial attention module that focuses on important spatial and structural features, and (3) a spectral preservation component that maintains spectral and radiometric integrity. The model incorporates components including dense blocks with spatial attentive enhancement, scale-adaptive modules for multi-scale feature extraction, channel and spatial attention mechanisms, and edge preservation modules. S³-ESRGAN was trained on our Sen2UAV dataset, comprising more than 428K low-resolution and high-resolution patch pairs from Sentinel-2 and corresponding UAV images across 33 scenes from diverse geographical regions across America, Europe, and Asia, covering various land cover types. The composite loss function integrates pixel-wise L1 loss, VGG-based perceptual loss, relativistic adversarial loss, spectral preservation loss, and edge preservation loss. Quantitative and qualitative image quality assessments demonstrate that S³-ESRGAN outperforms state-of-the-art methods such as Bicubic, SRCNN, VDSR, EDSR, SRResNet, and ESRGAN, achieving improved spectral fidelity with enhanced spatial details. Performance on our test set achieved the following results: PSNR: 20.89 dB, SSIM: 0.48, CC: 0.80, ERGAS: 9.77, SAM: 0.12 radians, EPI: 0.55, and SDI: 0.01. Spectral histogram analysis, spectral intercomparison, and point-wise evaluation collectively confirm the preservation of radiometric characteristics, with an average variation of 23% and an interband correlation coefficient of 0.79 when compared to the original and ground-truth images. Sub-pixel analysis using a slanted-edge target showed that the proposed method effectively reconstructs spatial features and improves effective spatial resolution by 42% compared to ESRGAN. This research contributes to the remote sensing community by enabling the generation of high-resolution imagery from open-access satellite data without costs.