Scholarlink

<p>Over the past few years, generative AI has captured the attention of researchers across various fields, demonstrating remarkable capabilities in creating synthetic data, particularly in image and video processing. Remote sensing, as the cornerstone of efficient earth resource monitoring and climate change assessment, stands to benefit significantly from these advances. However, certain remote sensing applications face critical limitations when dealing with regions characterized by small-holder agriculture and fragmented farmland. In such areas, the spatial resolution of freely available satellite imagery often proves insufficient for detailed analysis, creating a bottleneck in agricultural monitoring and management systems. To address this challenge, this study explores the application of Super-Resolution Generative Adversarial Networks (SRGAN) to enhance low-resolution Sentinel-2 images into high-quality PlanetScope imagery. Unlike super-resolution methods such as SRResNet, which focus primarily on pixel-wise reconstruction, SRGAN employs a more sophisticated approach by incorporating adversarial and perceptual losses alongside advanced techniques, including dense residual connections and multi-scale attention mechanisms. The experimental results demonstrate SRGAN’s superior performance across different land cover types. Urban areas achieved a Peak Signal-to-Noise Ratio (PSNR) of 33.15 and a Structural Similarity Index (SSIM) of 0.9030. Pure vegetation areas recorded an SSIM of 0.8782 and PSNR of 30.45, while water bodies showed a PSNR of 32.40 and SSIM of 0.9417. Barren land achieved a PSNR of 30.44 and an SSIM of 0.9109. These metrics significantly outperformed traditional methods (PSNR &lt; 11.5, SSIM &lt; 0.75), confirming SRGAN’s ability to preserve natural image characteristics while enhancing resolution. Validation against ground truth data revealed that the super-resolution approach achieved an overall accuracy of 89.3%, with consistent per-class performance: urban areas (89.74%), vegetation (88.69%), water bodies (91.00%), and barren land (91.30%). Overall, this study demonstrates that the application of SRGAN in remote sensing can significantly enhance image resolution and quality, which would help smallholder agriculture immensely by providing more detailed and accurate data for monitoring and management.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Novel Super-Resolution Residual GANs in Remote Sensing: A Case Study on Elevating Sentinel-2 to PlanetScope Imagery

  • 期刊论文

  • 发表日期：

    2026年5月14日

查看全文
  • Hadia Khan,
  • Nasru Minallah,
  • Waleed Khan,
  • Mahmood Ali Khan

摘要

Over the past few years, generative AI has captured the attention of researchers across various fields, demonstrating remarkable capabilities in creating synthetic data, particularly in image and video processing. Remote sensing, as the cornerstone of efficient earth resource monitoring and climate change assessment, stands to benefit significantly from these advances. However, certain remote sensing applications face critical limitations when dealing with regions characterized by small-holder agriculture and fragmented farmland. In such areas, the spatial resolution of freely available satellite imagery often proves insufficient for detailed analysis, creating a bottleneck in agricultural monitoring and management systems. To address this challenge, this study explores the application of Super-Resolution Generative Adversarial Networks (SRGAN) to enhance low-resolution Sentinel-2 images into high-quality PlanetScope imagery. Unlike super-resolution methods such as SRResNet, which focus primarily on pixel-wise reconstruction, SRGAN employs a more sophisticated approach by incorporating adversarial and perceptual losses alongside advanced techniques, including dense residual connections and multi-scale attention mechanisms. The experimental results demonstrate SRGAN’s superior performance across different land cover types. Urban areas achieved a Peak Signal-to-Noise Ratio (PSNR) of 33.15 and a Structural Similarity Index (SSIM) of 0.9030. Pure vegetation areas recorded an SSIM of 0.8782 and PSNR of 30.45, while water bodies showed a PSNR of 32.40 and SSIM of 0.9417. Barren land achieved a PSNR of 30.44 and an SSIM of 0.9109. These metrics significantly outperformed traditional methods (PSNR < 11.5, SSIM < 0.75), confirming SRGAN’s ability to preserve natural image characteristics while enhancing resolution. Validation against ground truth data revealed that the super-resolution approach achieved an overall accuracy of 89.3%, with consistent per-class performance: urban areas (89.74%), vegetation (88.69%), water bodies (91.00%), and barren land (91.30%). Overall, this study demonstrates that the application of SRGAN in remote sensing can significantly enhance image resolution and quality, which would help smallholder agriculture immensely by providing more detailed and accurate data for monitoring and management.