Fusion Of SAR And Optical Remote Sensing Image Based On Deep Convolutional Generative Adversarial Network

The fusion of SAR(Synthetic Aperture Radar)and optical remote sensing image has a wide range of application in scene classification,target detection,change detection and so on.However,it faces the challenges of insufficient effective data and short of feature extraction.The fusion of large-scale remote sensing images with high quality and computation performance based on deep learning and convolutional neural network,is of great value for multi-source satellite resources utilization and exploration of surface features information.In order to make full use of the texture information of SAR image and the color information of optical remote sensing image,this thesis propose a fusion method of SAR and optical remote sensing image based on the deep convolutional generative adversarial network.A method called SOGAN(SAR-Optical Generative Adversarial Network)is proposed to solve the fusion of medium-resolution remote sensing image.In this method,an encoder-decoder convolutional neural network is used as the generator of the fusion image,and a binary convolutional neural network is used to evaluate the quality of the fusion image.In order to learn the global and detailed features of the image,SO-GAN adopts a parallel feature extraction structure with multiple receptive fields.In addition,the loss function of the gray level co-occurrence matrix after HSV(Hue,Saturation,Value)transformation is added to constrain the texture characteristics of the image,so that the fused image can have the information of both SAR and optical image.Aiming at the task of high-resolution SAR and optical remote sensing image fusion,this thesis proposes an HRSO-GAN(High Resolution SAR-Optical Generative Adversarial Network)method.Based on multi-scale feature map pyramid generator and several discriminators.The image is generated layer by layer from low-resolution to highresolution.The more reasonable upsample operation is used to avoid the artifact of repetitive texture.And a multi-task loss function is introduced to further enrich the details.In this thesis,we use the SAR image obtained by sentinel-1A and the optical image obtained by sentinel-2 to carry out fusion experiments,use GF-2 optical image and GF-3 SAR image for high resolution image fusion.And use some objective image quality evaluation,such as P-SNR,to evaluate the color and texture quality of the fusion image.The experiment results show the validity and accuracy of our methods.