DADEMAS: Data Augmentation and DEnoising of MAgnetotelluric Signals based on CS-ResNet

Magnetotelluric (MT) sounding is a geophysical exploration method exploring the geoelectrical structure of the subsurface. However, weak MT signals are easily contaminated by various types of noise, which makes it difficult to acquire accurate information about the electrical properties and distribution characteristics of the subsurface rock layers. We develop a new MT denoising method for attenuating the noise in MT signals using data augmentation and CS-ResNet. First, we extract some noisy data samples from the measured MT data and combine them with samples constructed from mathematical functions to form an unexpanded sample set. Then, we apply the improved complete ensemble empirical mode decomposition with adaptive noise to obtain the intrinsic mode function (IMF) via the adaptive decomposition of the unexpanded training sample set. Next, we use the Pearson correlation coefficient to select the IMF and combine the selected IMF components with the unexpanded training sample set to form a training sample set for data augmentation. Finally, we input the data-augmented training set into the convolutional block attention module spatial pyramid pooling residual network (CS-ResNet) to learn the nonlinear mapping between the noise-containing data and noise. The experimental results of the simulated data and the measured data collected from the Luzong ore cluster area (117°25′E, 31°04′N), located in Anhui Province, China, indicate that this method can expand the diversity of the training sample set and better learn the characteristics of the MT noise data, making a significant improvement in the effectiveness and reliability of the denoising.