SAR Missing Echo Imaging via Hierarchical Learning Deployed on Hybrid Network

The fine quality of synthetic aperture radar (SAR) image made various kinds of downstream tasks feasible. Yet the received echoes suffered from the random missing circumstances in practical. In these conditions, the imaging quality was degraded, and hence deteriorated performance of downstream tasks. To address these problems, a learnable SAR imaging method was proposed in this paper. Different from the preceding works, more attentions were paid to generating high quality image with the random missing echoes. First, a filling and enhancing imaging network (FEINet) was developed, which is a hybrid network. The advantages of convolution and multi-head attention (MHA) mechanism were combined for SAR random missing echo imaging. Second, a hierarchical learning strategy was presented, which was called the deep supervision learning strategy. In this strategy, the filling imaging network and the enhancing imaging network were trained jointly. A joint loss measurement was then presented to guide the model to converge. Third, the converged model was used to generate the high-quality single-look complex (SLC) images from the random missing echoes. Multiple rounds of experiments were performed. The results proven that the proposed learning strategy played an important role for the performance of FEINet. And FEINet achieved the state-of-the-art (SOTA) imaging performance among all the compared methods.