Audio magnetotelluric denoising using time–frequency joint feature learning

The audio magnetotelluric (AMT) method is widely used in various mineral resource explorations due to its high‐frequency bandwidth and deep penetration capability. However, AMT signals are susceptible to cultural noise interference, which reduces the accuracy and reliability of AMT surveys. Existing denoising methods struggle to obtain high‐quality AMT data when dealing with strong cultural noise because they often ignore the time–frequency correlation features between the electric and magnetic channels. To address this issue, we propose a denoising method based on time–frequency joint feature learning. This method is based on a deep learning framework. During sample set creation, sliding time windows are used to extract electric and magnetic channel signals from the same time period, ensuring the correlation features between channels. Subsequently, wavelet time–frequency spectrograms are constructed as input to the neural network to deeply excavate the potential time–frequency characteristics of the noise, thereby improving denoising accuracy. Furthermore, we improve the U‐Net by adopting the exponential linear unit activation function and a hybrid pooling module, enabling joint capture of transient and steady‐state noise features. Through denoising experiments on synthetic AMT data and field AMT survey points, we compare our method with wavelet transform, data‐driven tight frame, traditional U‐Net and ResNet. Combining various evaluation metrics, such as normalized cross‐correlation, signal‐to‐noise ratio, reconstruction error, time–frequency analysis, apparent resistivity‐phase curves and Nyquist diagrams, we verify that our method can effectively enhance the quality of AMT data under strong cultural noise. Our method provides new insights and tools for high‐precision denoising technology in geophysical electromagnetic data processing.