Super-resolution Of Remote Sensing Images Based On Generative Adversarial Network Across Locations And Sensors

The accurate classification of land cover and land use provides an important data basis for the study of urban development and ecological environment,these tasks usually need high space-time resolution image.Due to the contradiction between the spatial resolution and the revisit period of remote sensing sensor,super resolution method is usually used to improve the spatial resolution of remote sensing image,and many learning-based models（neighborhood embedding,sparse representation,convolutional neural network,etc.）have been established.However,in the training and testing of these learning-based models,their data are often limited to a specific location and a single sensor,which results in the limited ability of the model to migrate across locations and sensors.In recent years,the generative model represented by the generative adversarial network GAN has shown great advantages in learning the distribution of samples and its high-dimensional characteristics,and the superresolution generative adversarial network（SRGAN）based on GAN has greatly improved the super-resolution effect.This paper research on the basis of SRGAN,in order to solve the training instability of SRGAN in the training firstly,the loss function of SRGAN is modified and the network structure is improved,we proposed improved SRGAN（ISRGAN）suitable for remote sensing image super-resolution,which made the model training more stable,enhanced the model transfer ability across locations and sensors,and realized the purpose of "one training is suitable for different regions and sensors" for remote sensing image.In the model transfer experiment,the training set and the test set were collected from Landsat 8 OLI and GF 1 remote sensing images in two different locations（Guangdong and Xinjiang）.In the cross-location universality test of the model,the model was trained with Guangdong GF 1 data and tested with Xinjiang GF 1 data.In the cross-sensor universality test of the model,the model trained in GF 1 in Guangdong was tested on Landsat 8 OLI images in Xinjiang.At the same time,the proposed method was compared with neighborhood embedding（NE）,sparse representation（SCSR）and SRGAN,and the peak signal-to-noise ratio（PSNR）and structural similarity（SSIM）were selected for quantitative evaluation.The results show that the proposed ISRGAN achieves an improvement of 3.8-5.6 on PSNR and 0.02-0.0.8 on SSIM compared with the other three super-resolution methods in the cross-location test of the model.In the cross-sensor test of the model,the proposed ISRGAN also achieves the improvement of 3.1-5.3 on PSNR and 0.01-0.04 on SSIM,respectively,and has the best superresolution performance.In addition,the end of this paper is application-oriented,taking land cover classification and land use classification as examples,we compared the accuracy of remote sensing images before and after super resolution based on ISRGAN in land cover classification and land use classification.The results show that the application of super resolution remote sensing image in land cover classification and land use classification significantly improves the classification accuracy,in which the accuracy of land cover classification is increased by 15% and land use classification is increased by about 50%.