Denoising framework for SFCW ground-penetrating radar under spectrally adjacent electromagnetic interference

Stepped-frequency continuous-wave ground-penetrating radar is susceptible to interference spectrally adjacent to its probing frequencies. After down-conversion, interference components with frequency offsets within the passband of the low-pass filter remain at nonzero baseband frequencies, producing background fluctuations that can obscure desired reflections. To address this problem, this study proposes a physics-guided denoising framework in a unified baseband domain, where desired echo responses are concentrated near zero frequency and spectrally adjacent interference is represented by frequency-offset components. Consecutive local batches are hierarchically organized according to interference intensity and morphology. Elite positive-region samples then guide class-level Bayesian-optimization variational mode decomposition (VMD), while boundary-aware adaptive reconstruction retains sample-level flexibility. Simulations under different interference levels demonstrate more accurate target-response reconstruction and better structural preservation compared with representative denoising methods. Reference-based validation confirms the reliability of the stability criterion, ablation analysis confirms the contributions of the principal framework components, and complexity profiling demonstrates the efficiency of class-level optimization. In field experiments, the mean amplitude of the deepest interface increases by 22.04% over averaging and 75.18% over fixed-parameter VMD. For anomalous-target detection, the proposed method consistently achieves the highest target-to-reference-region contrast while maintaining target-free background responses at levels comparable to or lower than those of the comparison methods. These results demonstrate reliable weak-response recovery and target-background separation under spectrally adjacent interference.