Deep Learning techniques for SAR data processing

Synthetic Aperture Radar (SAR) missions are experiencing a rapid increase in data volume due to higher spatial resolution, wider swaths, polarimetric acquisition and the deployment of large satellite constellations. This growth exacerbates well-known limitations in downlink capacity, onboard storage and ground-segment processing, making the traditional raw-data-centric pipeline progressively unsustainable. At the same time, many operational applications—such as rapid disaster mapping, maritime surveillance and deformation monitoring—require low-latency products that cannot rely exclusively on ground-based processing. This thesis investigates whether deep learning can provide an effective alternative to classical SAR focusing algorithms for onboard or near-sensor deployment. Specifically, it reformulates azimuth focusing as a one-dimensional sequence-to-sequence problem and proposes a state-space-model (SSM) architecture capable of operating in a fully streaming manner, avoiding the two-dimensional buffering and corner-turning operations that hinder conventional frequency-domain processors on embedded hardware. Experiments demonstrate that the distilled model achieves focusing quality comparable to the Range–Doppler Algorithm while offering substantial reductions in memory footprint and computational cost, making it a promising candidate for future onboard SAR processing. These results represent a step towards autonomous and adaptive cognitive SAR systems capable of real-time focusing and analytics directly in orbit.