Research On InSAR Phase Filtering Algorithm Based On Deep Learning

Interferometric Synthetic Aperture Radar(InSAR)is a quantitative microwave remote sensing measurement technology.Its basic principle is to coherently process the phase information of two time-domain SAR images in the same area.The interferometric phase is widely used in terrain mapping,surface deformation detection,surface change monitoring and atmospheric research.Since the quality of interference phase seriously affects the subsequent processing and practical application of InSAR data,it also determines the quality of elevation information generation.With the excellent performance of neural network in natural image denoising and super resolution,more and more researchers extend it to remote sensing image,medical image and astronomical image,and its effect is obviously superior to the traditional method.Influenced by the good performance of Neural networks,in this thesis,based on Denoising Convolutional Neural Network(Dn CNN),A novel Attention Mechanism Residual Learning-Interferometric Synthetic Aperture Network based on Attention Mechanism Residual Learning-Interferometric Synthetic Aperture Network,AR-InSAR Net).The main research contents are as follows:(1)InSAR technical process.Inorder to facilitate the development of the later research work,the principle of synthetic aperture radar and interferometry elevation information technology and InSAR processing process are introduced first.At the same time,the influence of noise on the subsequent processing and practical application of InSAR data is illustrated by computer simulation experiments.(2)Noise characteristics in interference phase diagram.The noise in the interference complex phase image is modeled,and the noise characteristics are analyzed,which provides a theoretical basis for the subsequent construction of neural network and data set.(3)Traditional InSAR phase graph filtering algorithm.The principles of different traditional InSAR phase graph denoising algorithms are described briefly,and the filtering performance and phase edge texture protection ability of these algorithms are verified by simulating phase image in interferometric complex domain.Inaddition,the results of different traditional InSAR phase graph filtering algorithms are analyzed quantitatively and qualitatively.The main innovations are as follows:(1)Interference phase filtering construction based on Dn CNN network.Inview of the excellent filtering performance of natural network,it is introduced into the field of InSAR interference phase filtering,and the interference phase filtering based on Dn CNN network is built.The specific filtering process of the network is as follows:Firstly,the interference phase of the real number domain is mapped to the complex domain,and then the phase of the complex domain is separated from the real part and the imaginary part through the vector separation technique,and then the separated real part and the imaginary part are input into Dn CNN respectively for filtering.Finally,the filtered phase image vector of the real part and the imaginary part is synthesized into the complex domain phase,and the filtered interference phase is synthesized through the angle function.Finally,the measured data generated by InSAR simulator is used to test the performance of the network.The experimental results show that the neural network is feasible to filter the interference phase.(2)AR-InSAR Net construction based on Dn CNN network.According to the noise model of phase diagram in interference-complex domain,based on the network framework of Dn CNN,the attention mechanism is introduced,the number of network channels is increased,and the AR-InSAR Net network is established,so that the effective phase information in phase diagram can be better protected in the process of filtering.At the same time,the constructed neural network is trained by using the constructed phase image data set.Finally,the simulated and measured interferometric phase diagrams are used to verify the constructed neural network.Through the subjective visual and objective indexes of the experimental results,the AR-InSAR Net filtering algorithm proposed in this thesis is obviously superior to the traditional filtering algorithm.