Audio-frequency magnetotellurics (AMT) is one of the commonly used methods in geophysical exploration; however, its signal energy is relatively weak and easily submerged by various cultural noises, making denoising a critical step in AMT data processing. Currently, deep learning-based neural networks have achieved superior denoising performance compared to traditional methods in many fields, but in AMT denoising, the neglect of the sparsity characteristic of cultural noise results in degraded denoising performance. To enable the neural network to consider sparsity features during the denoising process and thereby enhance denoising accuracy, we adopt a convolutional neural network (CNN) as the network backbone and design a multilevel wavelet convolutional neural network (MWCNN) from the perspective of sparse representation. This network improves CNN blocks via shortcut connections and enhances feature transmission efficiency by replacing pooling layers and interpolation with wavelet transforms, thereby enabling the network to account for the sparsity of cultural noise, capture underlying noise spectral information and improve denoising performance. Furthermore, we discuss the influence of various network parameters on denoising performance. Finally, we validate the effectiveness of MWCNN in AMT denoising through comparative experiments on both synthetic and field AMT datasets against wavelet transform, bounded influence remote reference processing, data-driven tight frame, CNN and residual networks. Comprehensive evaluations based on signal-to-noise ratio, wavelet time-frequency spectra, denoised results and residuals, apparent resistivity and phase curves, error analysis, one-dimensional inversion results and Nyquist diagrams confirm the superiority of MWCNN for AMT denoising.

Audio magnetotelluric (AMT) is commonly used in mineral resource exploration. However, the weak energy of AMT signals makes them susceptible to being overwhelmed by noise, leading to erroneous geophysical interpretations. In recent years, deep learning has been applied to AMT denoising and has shown better denoising performance compared to traditional methods. However, current deep learning denoising methods overlook the characteristics of AMT signals, resulting in reduced denoising accuracy. To enhance the denoising performance of deep learning by better matching the features of AMT signals, we propose a convolutional block attention module (CBAM)-based method for AMT denoising. This method focuses on the features of AMT signals and improves the process from three aspects: 1) in the establishment of the sample set, we adopt a multicomponent form based on the correlation of noise to enable the neural network to explore the potential connections among the components of AMT during the training process, thus constructing a stronger network mapping relationship; 2) in the construction of the neural network, we have introduced the CBAM structure into the residual blocks of the ResNet to enhance the network's feature learning capability by focusing on the characteristics of noise; and 3) in the design of the denoising procedure, we adopt a process of identification before denoising to protect the noise-free data segments from being compromised during the denoising process. Finally, through synthetic, field data experiments, and comparative tests, we demonstrate that our proposed method achieves higher denoising accuracy than some traditional methods and conventional deep learning methods.