GPU-based TOPS Mode Data Interferometric Registration And Filtering

Sentinel satellite provides a kind of Interferometric Synthetic Aperture Radar(In SAR)data,which can be used to extract ground elevation and deformation information.It has a wide range of uses.The Terrain Observation By Progressive Scans(TOPS)working mode requires registration accuracy better than one-thousandth of a pixel,which brings a huge challenge to image registration;the sentinel data has the characteristics of high resolution and continuous spatial coverage.At the same time,the large amount of image data and the long time-consuming registration bring new challenges to fast data processing.Interferogram filtering is a key step in interference processing.However,the existing filtering methods cannot retain the features in the dense fringes.The phase estimation accuracy of the coherent region is low.In response to the above problems,this paper studies the high-precision registration of TOPS mode and how to accelerate it,and introduces Convolutional Neural Networks(CNN)to learn noise features to improve the interferogram filtering effect.The specific content is as follows:(1)In this paper,I analyze the imaging principle of Sentinel-1A TOPS mode and conduct research on its high-precision registration.First select the external DEM data and the precise orbit data to perform geometric coarse registration of the image,then use the coherence coefficient method based on the signal-to-noise ratio for the second-level coarse registration,and finally use Enhanced Spectral Diversity(ESD).The experimental results show that the final azimuth registration accuracy is0.000445,which is better than one-thousandth of a pixel,it meets the registration accuracy requirements of TOPS mode images.(2)In order to solve the time-consuming problem of TOPS mode images,MPI(Message Passing Interface)and CUDA(Computing Unified Device Architecture)are used to respectively accelerate geometric registration,coherence coefficient method registration,and enhanced spectral diversity registration.Experimental verification shows that compared with a single-threaded CPU(Central Processing Unit),the six-threaded CPU speedup ratio reaches 3.32,and the GPU(Graphics Processing Unit)speedup ratio can reach 112.75.(3)An interferogram filtering method based on convolutional neural network is proposed.This method uses the self-encoder structure for unsupervised learning,takes the residual noise after removing the slope phase of the local terrain as the model input,and then adds the filtering result to the phase of the local terrain slope.This method can maximize the recovery of interferogram image information from the noisy interferogram.Experiments with simulated data and real data are compared with Goldstein filtering,mean filtering,Lee filtering,Frost filtering,and improved Dn CNN.The results show that this method can greatly improve the phase quality of the interferogram and suppress noise,and at the same time,it can restore more image details and maintain the continuity of interference fringe edges.