Deep learning for SAR focusing

Image focusing is a key intermediate step in the larger SAR system pipeline for various applications. While improving the numerical accuracy of SAR image focusing algorithms is possible through computational trade-offs, the visual quality of the scene is seldom the desired outcome in a SAR image. Applications such as object detection, segmentation, or target classification each have a different objective and thus a different notion of quality. However, regardless of the objective, all SAR systems share limitations, as system uncertainties can have devastating effects that can be traced to the degradation in the focused SAR image. For example, noisy and corrupted measurements, poor collection geometries, finite instantaneous bandwidth, and other uncertainties of the collection system will deteriorate image quality independent of the algorithm or the downstream tasks of the SAR system. In this chapter, we discuss a framework for applying deep learning to SAR image formation developed from first principles to directly address system uncertainties, which are demonstrated with applications to SAR autofocus, and passive SAR for dealing with phase errors.