Image Super-Resolution Reconstruction of Landslide Based on Real-ESRGAN

The detection and extraction of landslide remote sensing images is of great significance for disaster monitoring and forewarning, risk assessments, and disaster prevention and control. However, the collected remote sensing images are not clear enough due to equipment limitations, which affects the application of remote sensing interpretation. For the lack of high-quality landslide datasets in the process of remote sensing images interpretation, we propose a model based on Real-ESRGAN to achieve super-resolution reconstruction of low-resolution landslide images in this paper. The method we proposed embeds a second-order degradation process combined with random shuffling strategy into the ESRGAN model to simulate the degradation process of the real images, choosing a U-Net discriminator with spectrum normalization to increase the discrimination ability and stabilize training dynamics. With transfer learning algorithm, experimental verification was carried out based on aerial images of landslides around Sichuan. The experimental results show that the Real-ESRGAN model based on transfer learning achieves higher scores in peak signal-to-noise ratio (PSNR) and structural similarity (SSIM), and achieves better results in super-resolution reconstruction.