Deep learning-based pseudo-focusing of SAR images for onboard ap- plications

Abstract

Real-time Synthetic Aperture Radar (SAR) processing on board satellites is essential for enabling autonomous, low-latency Earth Observation (EO) missions. Yet, traditional SAR focusing techniques remain highly computationally demanding for execution on resource-limited space platforms. To address this challenge, we propose a physics-informed deep learning approach that performs pseudo-focusing of SAR amplitude images. Our method employs an ad hoc-built convolutional neural network (CNN) trained to emulate the output of conventional focusing algorithms while drastically reducing processing time and computational load. The resulting pseudo-focused images preserve key semantic content and demonstrate sufficient fidelity for downstream onboard tasks powered by machine learning. Experimental evaluations confirm that this approach can deliver near-real-time SAR imaging performance, paving the way for intelligent, adaptive SAR payloads in future EO missions.