A Robust RFI Mitigation Method for SAIR Using a Multiscale Dilated and Detail Enhancement Network

Abstract

Radio frequency interference (RFI) in the L-band remains a challenge for synthetic aperture interferometric radiometers (SAIRs), degrading data quality and brightness temperature (BT) retrieval accuracy. To address this issue, we propose the multiscale dilated and detail enhancement network (MDDEN), a deep learning (DL)-based method that directly suppresses RFI in the BT domain without requiring auxiliary prior knowledge. MDDEN integrates a multidilated module (MDM) to capture spatially complex interference patterns and a detail enhancement module (DEM) to recover high-frequency BT information. Furthermore, we propose the cross correlation reconstruction error (CCRE), a no-reference metric designed to evaluate RFI mitigation performance. Experiments on both simulated and observed soil moisture and ocean salinity (SMOS) data demonstrate that MDDEN achieves robust and generalizable mitigation across diverse scenarios, delivering high-fidelity BT reconstructions and offering a practical solution for SAIR data processing under RFI contamination.