Deep SAR Imaging and Motion Compensation.

Abstract

Compressive sensing (CS) and matrix sensing (MS) techniques have been applied to the synthetic aperture radar (SAR) imaging problem to reduce the sampling amount of SAR echo using the sparse or low-rank prior information. To further exploit the redundancy and improve sampling efficiency, we take a different approach, wherein a deep SAR imaging algorithm is proposed. The main idea is to exploit the redundancy of the backscattering coefficient using an auto-encoder structure, wherein the hidden latent layer in auto-encoder has lower dimension and less parameters than the backscattering coefficient layer. Based on the auto-encoder model, the parameters of the auto-encoder structure and the backscattering coefficient are estimated simultaneously by optimizing the reconstruction loss associated with the down-sampled SAR echo. In addition, in order to meet the practical application requirements, a deep SAR motion compensation algorithm is proposed to eliminate the effect of motion errors on imaging results. The effectiveness of the proposed algorithms is verified on both simulated and real SAR data.