Refocusing of Moving Ships Based on Deep Learning for High Altitude Platform SAR

Imaging maritime vessels in synthetic aperture radar (SAR) systems has long been a challenging task. The translational and six-degree-of-freedom complex motions of vessels introduce imaging artifacts such as shifts and defocusing. Conventional approaches typically require estimation of motion parameters to refocus. In contrast, deep learning methods cast the refocusing problem into a regression framework. However, existing deep learning refocusing approaches mostly address short aperture times and are tailored to airborne or spaceborne platforms where vessel-induced shifts and defocusing are relatively minor. However, for slow-moving high-altitude platforms (HAP) with extended aperture times, these approaches become less effective. Hence, this paper derives an analytical expression for the signal model in the image domain after pulse compression for maritime vessels with six-degree-of-freedom motion. Various parameters affecting the complex motion are analyzed, and the derived model is utilized to generate subsequent training set images. Subsequently, a refocusing approach based on Resnet50 is proposed, incorporating a feature fusion module that utilizes the generated images from the analytical expression for training. Finally, through comparative analysis with other networks, the proposed approach demonstrates superior refocusing performance for SAR images with significant defocusing under long apertures in slow-moving, high-altitude platforms.