Deep learning for onboard SAR intelligent applications

Kurzfassung

Onboard Synthetic Aperture Radar (SAR) processing is a key enabler for low-latency, autonomous Earth Observation (EO) missions, particularly within the frame of cognitive radars. However, conventional SAR focusing techniques are computationally demanding and difficult to optimize for real-time execution on resource-constrained platforms, thereby limiting the deployment of advanced onboard applications such as target detection, situational awareness, and real-time monitoring. This paper introduces a physics-informed, deep learning-based framework, named FocusNet, for fast focusing of SAR images. The proposed method leverages a convolutional neural network specifically designed to approximate the output of conventional focusing algorithms, with both the network architecture and training strategy grounded in the theoretical principles of SAR image formation. Moreover, the framework is conceived with the long-term objective of enabling a fully integrated, end-to-end AI-driven pipeline for onboard SAR applications, operating directly on raw data without relying on the traditional two-stage processing chain consisting of a separate focusing step followed by higher-level analysis. This paradigm has the potential to streamline processing, reduce computational overhead, and accelerate inference. Experimental results demonstrate that the reconstructed SAR images preserve critical semantic content and reliably reproduce characteristic SAR spatial patterns. These findings underscore the promise of deep learning-based SAR processing as a foundation for next-generation intelligent SAR payloads.