Cost-Aware Uncertainty-Gated Cascade for Onboard RFI Detection in SAR Raw Echoes
Abstract
Radio frequency interference (RFI) detection is an important front-end step for synthetic aperture radar (SAR) interference mitigation. To meet the robustness and onboard real-time processing requirements of SAR RFI detection, this letter proposes a cost-aware uncertainty-gated cascade method for pulse-level RFI detection in SAR raw echoes. The method invokes time-domain, frequency-domain, and higher-order features in order of increasing computational cost, and generates confidence scores at each stage. High-confidence pulses are decided early, whereas uncertain pulses are sent to the next stage. The first stage performs fast time-domain screening using local range-block energy deviations. The second stage constructs a frequency-domain reference from first-stage high-confidence RFI-free pulses and computes frequency-band anomaly confidence for the remaining pulses. The third stage extracts time–frequency features and lightweight subspace features only for the remaining uncertain pulses, and trains Bagged Trees with confidence-weighted pseudo labels for the final decision. Experiments show that the proposed method achieves a 93.3% detection rate at SIR = 3 dB. On Sentinel-1A measured data, the first two stages decide 88.38% of valid pulses in advance with 98.46% accuracy, and the complete cascade achieves 97.60% accuracy and a 1.02% false alarm rate while maintaining second-scale processing per burst.