基于生成对抗网络的Sentinel-2遥感图像超分辨率分析

Super-resolution analysis of Sentinel-2 images using generative adversarial networks

为了将开放访问的Sentinel-2卫星遥感图像的分辨率提升至商业卫星的水平,提出基于生成对抗网络(GAN)的超分辨率分析方法KN-SRGAN,该方法仅使用开放数据提供的图像,不须高分辨率监督图像,通过核估计和噪声注入构造高-低分辨率图像对训练数据集,构建带有感知特征提取器的GAN,实现卫星图像×4倍的超分辨率分析。与残差通道注意力网络(RCAN)、强化深度残差网络(EDSR)、强化超分辨率生成对抗网络(ESRGAN)、退化核超分辨率生成对抗网络(DKN-SR-GAN)等最新方法比较,KN-SRGAN的生成图在直观视觉效果上具有更清晰的细节以及更好的感知效果,无参考图像质量评估指标的定量对比也证明了KN-SRGAN的有效性。

To improve the resolution of open access Sentinel-2 remoted sensing iimages to the level of commercial satellites,a super-resolution method based on a Generative Adversarial Network(GAN)was proposed,namely KN-SRGAN,which only used the images provided by open data without high-resolution supervision images,constructed a high-low resolution pairs’training dataset by kernel estimation and noise injection,and established a GAN with perceptual feature extractor to realize×4 times super-resolution analysis of satellite images.Compared with the latest methods such as Residual Channel Attention Network(RCAN),Enhanced Deep Residual Network(EDSR),Enhanced Super Resolution Generative Adversarial Network(ESRGAN)and Degraded Kernel Super Resolution Generative Adversarial Network(DKN-SR-GAN),the generated image of KN-SRGAN has more precise details and better perception effect in visual effect,and the quantitative comparison of Non-reference Image Quality Assessment(NR-IQA)also proves the effectiveness of KN-SRGAN.