Super-resolution Method Of Remote Sensing Image Based On Generative Adversarial Network

Remote sensing image is an important data source to quickly obtain large range of ground information,which plays an important role in many fields such as agricultural monitoring,resource and environment investigation.Therefore,the acquisition of high spatial resolution remote sensing images is of great significance to academic study and production practice.Super-resolution is a technology that converts low-resolution images into high-resolution images to improve the spatial resolution of images.It is widely used in video surveillance and medical images.With the development of deep learning technology,super-resolution models based on a variety of neural network algorithms have been developed,among which the use of generative adversarial network to obtain high-resolution images is a hot topic of study at present.But the network is mostly for camera image applications.Remote sensing images contain complex features of ground objects,diverse types of covered ground objects,and rich spectral bands.In the results obtained by using the existing generative adversarial network,there are some problems such as rough texture details and unreal color distribution.To solve these problems,a Remote Sensing Super Resolution Generative Adversarial Network（RSSRGAN）is proposed in this dissertation.The main work of this dissertation is as follows:1.Making data sets.In this dissertation,17 GF1 wide-field multispectral images and 56GF2 multispectral images obtained from January 2019 to December 2020 in Shandong and Ningxia were preprocessed by radiometric calibration,atmospheric correction,orthographic correction and geographical registration.The GF1-GF2 data set was made by using the preprocessed image,with Gaofen-1 wide-view-field multi-spectral image as the low-resolution image and Gaofen-2 multi-spectral image as the reference image.In addition,in order to further test the performance of the model,the Potsdam dataset is also made in this dissertation.2.RSSRGAN model construction.RSSRGAN is composed of generator and discriminator.In terms of generator,the existing convolutional neural network architecture is analyzed,and a network architecture that combines residual structure with multi-layer feature fusion strategy is proposed.This architecture extracts multi-level feature information through residual structure.Multi-level feature fusion strategy is used to fuse multi-level high-level features and low-level features.Enhance the simulation ability of the generator network to improve the quality of generated images.In order to ensure the realistic color of the generated image,a filter block modeled in the feature channel is designed to constrain the numerical distribution range of the features in the image reconstruction process of the generator.Further enhance the consistency of color distribution between low-resolution image and reference image.In terms of discriminator,the discriminator method of image pixel by pixel prediction is adopted.The discriminator network of encoder-decoder architecture is designed according to this method,and the prediction result image which can describe the probability of all pixels of the image is obtained.Enhance the discriminator’s perception ability to image details.3.Analyze the results of experiment.The traditional Bicubic interpolation,EDSR,SRGAN and ESRGAN models are used as comparison models to carry out comparative experiments with the RSSRGAN model.The experimental results show that the SSIM（Structural SIMilarity）performance metrics of RSSRGAN exceeds the experimental comparison model,reaching 0.955,indicating its superiority in super resolution of remote sensing images.In addition,this dissertation set up three modules of multi-layer feature fusion strategy,filter block and discriminator per-pixel prediction validity test.The experimental results show that the combination of the three modules in the proposed method can further improve the accuracy of the generated image.In this dissertation,RSSRGAN model is constructed for the characteristics of remote sensing images.This method has an excellent effect in improving the spatial resolution of remote sensing images.It can be used to obtain remote sensing images with high spatial resolution and large coverage area.It is of great significance to expand the application range of domestic remote sensing data,expand the data sources of domestic satellite images,and improve the precision of mapping the spatial distribution of ground objects covering a large area.