Self-supervised deep learning approach for SAR narrowband RFI suppression using range frequency sub-band decomposition

Radio Frequency Interference (RFI) is a growing challenge in Synthetic Aperture Radar (SAR) Imagery, where signals from ground-based emitters cause artifacts that degrade image quality, obscure areas of interest and negatively impact the performance of automated SAR analysis methods.

This work proposes a self-supervised deep learning approach for RFI suppression in Sentinel-1 IW Single Look Complex (SLC) data, requiring no clean reference imagery or data labeling and therefore overcoming the limited availability of annotated complex SAR datasets for supervised learning. This approach exploits the spectral characteristics of narrowband RFI, which occupies a small fraction of the total SAR bandwidth. When an SLC image is decomposed into a stack of range frequency sub-bands, the interference appears at a different range position in each sub-band, revealing areas that were masked by the RFI in the full bandwidth image. This inconsistent positioning across sub-bands is used as an automatic supervision to train a U-net architecture in a Noise2Self-inspired framework, aimed at reconstructing a clean full-resolution image from a stack of range frequency sub-bands.

Preliminary analysis confirms the spectral shift behavior of NBI across sub-bands, supporting the feasibility of the proposed supervision strategy. Initial training results show measurable RFI suppression, with performance depending on the fraction of the spectrum occupied by the interference. RFI confined to a small portion of the bandwidth is easily suppressed, while interference occupying a larger fraction is partially suppressed.