Adaptive Parameter Selection of SAR Sparse Imaging Model
SUN Dou,XING Shiqi,LI Yongzhen,DAI Dahai
DOI: https://doi.org/10.11834/jrs.20176207
2017-01-01
Abstract:High resolution is an important direction for the development of Synthetic Aperture Radar (SAR).Conventional SAR imaging algorithms based on matched filtering is limited by the signal bandwidth and the synthetic aperture length.Compressive Sensing (CS) theory has recently attracted the attention of domestic and foreign scholars.Electromagnetic scattering signals of man-made objects,such as vehicles and buildings,have obvious sparse characteristics,and the total electromagnetic scattering of the target can be approximated by the synthesis of the center scattering of the local scattering.Studies have shown that CS has considerable potential in improving the quality of radar imaging and high precision imaging of man-made targets.However,the application of CS theory in radar imaging is novel;therefore,it has some problems that require further study.Studies have shown that the regularization and shrinkage parameter have an obvious impact on the performance of radar sparse imaging.Traditional parameter selection methods have serious disadvantages in computation cost and accuracy or are limited to a few specific models.Therefore,proposing a parameter selection method is of great importance.A fully automated algorithm is proposed in this paper to improve the performance of radar sparse imaging in the case of parameter optimization and to overcome the preceding disadvantages mentioned.First,combined with the Maximum A Posteriori (MAP) estimation and Bayesian inference,we obtain the relationship between regularization and shrinkage parameter.The value of regularization parameter is determined by the choices of shrinkage parameter,noise,and signal variance,thereby reducing the dimension of the parameter selection problem.Second,we propose a radar sparse imaging algorithm which can solve the joint optimization problem to simultaneously achieve model parameters estimation and SAR imaging.In each iterative of the approach,all required parameters are updated with training data without the necessary prior information,and the image is reconstructed by regularization technique with the updated observation model constructed by new parameters.Extensive comparisons are carried out between the proposed method and several other competing methods based on simulations and real-data processes.The experimental results demonstrate that the proposed method can achieve accurate parameter estimation and imaging performance with low computational complexity.Compared with the traditional imaging method,the imaging results of this method are sparse,thereby reducing the sidelobes and maintaining cleanliness without noise.This paper presents a fully automated parameter selection method based on maximum a posteriori estimation and Bayesian inference without extra prior information.All required parameters can be obtained through known data.We deduce the relationship between model parameters and signal and noise variance;thus,a series of training processes has been avoided,and the computational cost has been considerably reduced.Simulations and realdata experiment demonstrate that the proposed method can achieve accurate parameter selection with lower computational cost than other parameter selection methods,such as Bayesian information criteria and L-curve.