Time–frequency collaborative denoising for audio-magnetotelluric data using a wavelet-based residual network

Abstract The Audio-Magnetotelluric (AMT) method is a key geophysical technique for mineral resource exploration, but anthropogenic electromagnetic interference severely downgrades the data quality. In recent years, neural networks have shown results superior to traditional methods for AMT time-domain denoising. However, existing approaches often overlook deep-seated signal characteristics, leading to suboptimal performance in processing low-frequency data. To address such limitations, we introduce an innovative time-frequency collaborative network—Wavelet-Based Residual Network (WaveResNet). Distinct from conventional single-domain (time/frequency) processing techniques, WaveResNet incorporates a tailored wavelet convolutional architecture that effectively integrates temporal and spectral attributes of AMT signals. By concatenating features from wavelet-decomposed subcomponents and enabling collaborative learning, the network profoundly exploits coupled time-frequency signatures, notably enhancing separation capability for complex anthropogenic noise. Concurrently, the downsampling effect inherent to wavelet decomposition effectively mitigates processing loss in meaningful signals. Furthermore, WaveResNet synchronously models all four electromagnetic field components, fully leveraging inter-channel correlations. The proposed workflow follows a "Detect-and-Denoise" strategy, where only noisy segments are processed, thereby preserving the integrity of low-noise data. Experiments on both synthetic and field data demonstrate the method effectively identifies and suppresses AMT noise, outperforming existing network-based approaches and offering a novel solution for high-fidelity denoising in environments with strong interference.