基于深度展开的SAR大斜视RD成像算法

大斜视角条件下的合成孔径雷达SAR回波信号具有方位向和距离向严重耦合、大距离徙动等特点，采用常规的距离多普勒RD算法成像，会引起方位向的散焦以及空变等问题。为了改善大斜视SAR在成像过程中存在的问题，提出了一种基于深度展开网络的SAR大斜视可学习距离多普勒成像方法。该方法将RD成像方法与深度学习结合，利用RD成像的步骤构建了基于深度展开网络的RD学习成像网络结构，将回波数据作为网络输入来学习回波数据到大斜视SAR图像的成像过程。首先，在分析大斜视SAR回波信号模型的基础上确定了网络成像过程中的可学习参数；其次，根据成像过程设计大斜视SAR成像网络；最后，通过非监督训练的方法对网络进行训练，最终输出学习成像结果。点目标和面目标仿真结果表明，该方法可以有效抑制旁瓣，提高成像精度和计算效率，满足SAR在大斜视角下的成像要求。

A Novel Range-Doppler Imaging Method for Highly Squinted SAR Based on Deep Unfolded Net

The high-squint synthetic aperture radar (SAR) echo signal is characterized by serious coupling between azimuth di-rection and range direction, large range migration. Imaging with conventional range Doppler (RD) algorithm causes problems such as azimuth defocus and space variation. In order to solve the problems in the imaging process of high-squint SAR and imaging quality and calculation cost, a Learnable Range Doppler (LRD) imaging method for high-squint SAR based on deep unfolded network is proposed. This method combines RD algorithm with deep learning by using RD imaging depth steps to build an RD imaging study network structure. Taking the echo data as network input is to learn the imaging process from echo data to SAR image. Firstly, the learnable parameters of imaging network are determined based on the analysis of the echo signal model of high-squint SAR. Secondly, the SAR imaging network is designed according to the imaging process. Finally, the network is trained through unsupervised training method and output the learning imaging result. The simulation results of point targets and real scene targets show that the proposed method can effectively suppress side lobes, improve imaging accuracy and calculation efficiency, and meet the requirements of high-squint SAR imaging.