Abstract:
The magnetotelluric (MT) method is a core technology widely used in geophysical exploration and commonly used in geological surveys, resource exploration and geodynamic research. However, MT data are susceptible to complex noise, including nonlinear and non-stationary noise, which significantly reduces data quality and interpretation accuracy. Although traditional denoising methods (such as sparse representation and wavelet transform) can improve the quality of some data, they have limitations such as complex parameter adjustment and insufficient robustness when dealing with diverse noise types. In order to solve the above problems, this paper proposes an innovative MT data denoising method based on inverse autoencoder and channel attention mechanism. The inverse autoencoder enhances the ability to capture complex signal features through the process of dimensionality increase and dimensionality reduction, achieving efficient signal-to-noise identification and signal fitting; the channel attention mechanism further improves denoising accuracy by dynamically adjusting the weight of feature channels. On this basis, an end-to-end deep learning framework is designed to process MT data in complex noisy environments. Experimental results show that this method exhibits superior denoising performance under a variety of noise conditions. It is significantly better than traditional methods in indicators such as correlation coefficient (CORC), normalized root mean square error (NRMSE) and signal-to-noise ratio (SNR); in addition, in the analysis of apparent resistivity-phase curve and electromagnetic field polarization direction , the method in this paper shows higher robustness and consistency. This shows that the method in this paper can effectively improve the quality and interpretability of MT data and provide reliable technical support for geophysical exploration.