Towards end-to-end SAR raw data handling: RCMC data compression using hyperprior autoencoders

Kurzfassung

Synthetic Aperture Radar (SAR) is a fundamental imaging technique in Remote Sensing. While modern SAR systems offer unprecedented resolution and coverage, they generate large volumes of raw data that cannot be processed directly onboard and must be transmitted to the ground station, straining downlink bandwidth. In an effort to bring efficient SAR data compression onboard, this paper explores the possibility of compressing intermediate representations of SAR data. In particular, we experiment with the compression of Range Cell Migration Corrected (RCMC) data using hyperprior autoencoders. To this end, we adapt existing deep learning compression models to the specific characteristics of SAR RCMC data and introduce a task-aware training strategy that enforces coherence preservation. Experimental results show that hyperprior models achieve up to 9dB PSNR improvement over traditional codecs, while preserving complex coherence up to 0.95 at high bitrates.