Nonstationary SBL for SAR Imagery Under Nonzero Mean Additive Interference

Abstract

Conventional sparse Bayesian learning (SBL) for synthetic aperture radar (SAR) imagery predominantly focuses on the signal model which is affected by the zero mean white Gaussian noise. Therefore, it is difficult to solve the problem of non-zero mean additive interference. In this letter, a new and non-stationary SBL (NS-SBL) algorithm is proposed, which is capable of handling the additive interference with non-zero mean and in general distributions. Specifically, to accommodate the strong direct-current power of the interference, the Gaussian mixture model (GMM) is introduced to fit the expectation, which is flexible enough to capture the complicated variation of additive interference. However, the intended additive perturbation will result in complicated likelihood distribution, which is difficult to get a fully Bayesian inference. To this end, hierarchical Bayesian inference is employed to solve the problem that the likelihood and prior are not conjugated. Further, the variational Bayesian (VB) method is adopted to overcome the difficulties in accessing closed-form solutions of the intended posteriors. Experimental results of synthetic and practical data indicate that the proposed algorithm has superior performance of anti-interference and super-resolution over other reported ones, especially under limited data and complicated environments.