Abstract: To address the problem of the lack of high-resolution training set in the use of high-resolution remote sensing images for landslide intelligent recognition. In the process of forming the high-resolution landslide training set, in order to make full use of the previous low-resolution landslide images, this paper adopts enhanced super-resolution generative adversarial networks(ESRGAN) to achieve super-resolution reconstruction of the low-resolution landslide image set. The ESRGAN model improves the feature extraction performance and stability of the generator by removing the batch normalization layer, adding multi-stage residual network and residual scaling factor based on the SRGAN model, and using the transfer learning method to conduct experimental based on the Bijie landslide image set and Yunnan Nanjing highway landslide image set. The experimental results show that the ESRGAN model based on transfer learning can achieve higher scores in peak signal-to-noise ratio (PSNR) and structural similarity (SSIM), and better results in super-resolution reconstruction. The results of this paper can provide a new technical method for acquiring high-resolution remote sensing image sets of landslides.

Key words: landslide image set, adversarial network generation, super-resolution reconstruction, multistage residual network, transfer learning