When the controlled-source electromagnetic (CSEM) data are contaminated by intense cultural noise and the signal-to-noise ratio (S/N) is lower than 0 dB, the existing denoising methods can hardly achieve good results. To overcome the problem, a new strong-noise elimination method called inception-temporal convolutional network-shift-invariant sparse coding (IncepTCN-SISC) is developed based on deep learning and dictionary learning. First, a novel deep neural network model called IncepTCN is created based on the inception block and temporal convolutional network (TCN). Then, IncepTCN is used to recognize strong-noise segments in the observed signal, which are then discarded. Finally, a dictionary-learning method based on shift-invariant convolutional coding is used to denoise the remaining weak-noise segments. A series of simulated and field data experiments indicate that the new proposed IncepTCN network has obvious advantages in accuracy and efficiency compared with alternative methods. The average recognition accuracy of IncepTCN is 96.5%, which is 25.5%, 3.2%, 1.1%, and 2.0% higher than that of the fuzzy C-means clustering, convolutional neural network (CNN), residual network (ResNet), and the nonimproved TCN, respectively. In addition, the test results of unfamiliar data indicate that the generalization ability of IncepTCN is significantly better than the CNN, ResNet, and nonimproved TCN. This IncepTCN-SISC method can improve the S/N of CSEM data from −5.0 dB to 3.1 dB or from 5.0 dB to 31.9 dB and solve the denoising problem of noisy data below 0 dB to a certain extent. After IncepTCN-SISC processing, the initially distorted apparent resistivity curves become smooth, and the result is better than dictionary learning. This method is intelligent without any manual intervention and is suitable for batch processing of CSEM data.

Magnetotelluric (MT) method is widely used for revealing deep electrical structure. However, natural MT signals are susceptible to cultural noises. In particular, the existing data-processing methods usually fail to work when MT data are contaminated by persistent or coherent noises. To improve the quality of MT data collected with strong ambient noises, we propose a novel time-series editing method based on the improved shift-invariant sparse coding (ISISC), a data-driven machine learning algorithm. First, a redundant dictionary is learned autonomously from the raw MT data. Second, cultural noises are reconstructed using the learned dictionary and the orthogonal matching pursuit (OMP) algorithm. Finally, the de-noised MT data are obtained by subtracting the reconstructed cultural noises from the raw MT data. The synthetic data, field experimental data and measured data are tested to verify the effectiveness of the newly proposed method. The results show that our new scheme can effectively remove strong cultural noises and has better adaptability and efficiency than the predefined dictionary-based methods. The method can be used as an alternative when a remote reference station is not available.

Controlled-source electromagnetic (CSEM) data recorded in industrialized areas are inevitably contaminated by strong cultural noise. Traditional noise attenuation methods are often ineffective for intricate aperiodic noise. To address the abovementioned problem, we have developed a novel noise isolation method based on the fast Fourier transform, complementary ensemble empirical mode decomposition (CEEMD), and shift-invariant sparse coding (SISC, an unsupervised machine-learning algorithm under a data-driven framework). First, large powerline noise is accurately subtracted in the frequency domain. Then, the CEEMD-based algorithm is used to correct the large baseline drift. Finally, taking advantage of the sparsity of periodic signals, SISC is applied to autonomously learn a feature atom (the useful signal with a length of one period) from the detrended signal and recover the CSEM signal with high accuracy. We determine the performance of the SISC by comparing it with three other promising signal processing methods, such as the mathematic morphology filtering, soft-threshold wavelet filtering, and K-singular-value decomposition (another dictionary learning method) sparse decomposition. Experimental results illustrate that SISC provides the best performance. Robustness test results indicate that SISC can increase the signal-to-noise ratio of noisy signal from 0 to more than 15 dB. Case studies of synthetic and real data collected in the Chinese provinces of Sichuan and Yunnan indicate that our method is capable of effectively recovering the useful signal from the observed data contaminated with different kinds of strong ambient noise. The curves of U/I and apparent resistivity after applying our method improved greatly. Moreover, our method performs better than the robust estimation method based on correlation analysis.

The magnetotelluric (MT) signals are susceptible to anthropogenic noise and the existing denoising methods have significant shortcomings in low-frequency situations. To address the problem, we propose an innovative denoising approach. It is different from the existing methods that attempt to achieve signal-noise separation through one step. The denoising process is divided into two steps in the proposed approach. The effective low-frequency dominant component and high-frequency component are sequentially extracted through deep learning and dictionary learning. We propose a new deep learning network named DnCNN-GRU which combines the powerful feature extraction capability of Denoising Convolutional Neural Network (DnCNN) and the strong temporal sequence processing ability of Gated Recurrent Unit (GRU), enabling accurate extraction of the low-frequency MT signal. Furthermore, we integrate this network with the K-Singular Value Decomposition (KSVD) dictionary learning to achieve accurately extraction of effective high-frequency components. Tests of synthetic data indicate that our method is the best compared to a series of state-of-the-art (SOTA) algorithms. It is the only method that can completely remove various types and scales of cultural noises while brilliantly preserves both the low and high-frequency signals. In addition, our method is validated on apparent resistivity and phase data and is significantly superior to the commonly used Robust estimation method. These results demonstrate that our method can solve the problem mentioned above and can be a substitute for Robust estimation or remote reference processing.

Geomagnetic data are widely used in earthquake prediction, mantle conductivity imaging, and other fields. However, the problem of geomagnetic data being contaminated by cultural noise is becoming increasingly serious. Existing denoising methods have shortcomings such as insufficient flexibility and the need for manual intervention. To this end, we modify the U-net and propose a new intelligent geomagnetic signal denoising method based on the network. The novel network not only combines the advantages of denoising convolutional neural network (DnCNN) and U-net, but also utilizes the shortcut connections to prevent network degradation. We obtain a high-precision denoising model through elaborate training sets. The processing results of synthetic data show that the improved U-Net can remove various types of noise in one step, such as impulse noise, square wave noise, and Gaussian noise. The signal-to-noise ratio (SNR) of the denoised signal increases by an average of more than 20 dB, and the average normalized-cross correlation (NCC) between the denoised signal and the high-quality signal reaches 0.9998. Compared with Wavelet threshold denoising, DnCNN, and U-Net, the improved U-Net has obvious advantages. We apply the method to real geomagnetic data collected in Guangxi, Yunnan, Gansu, and Tibet, China. The results demonstrate that the proposed method can significantly improve the tippers, coherencies, and induction arrows. Compared with traditional methods, our method eliminates subjective bias, improves the adaptability to different types of noise, and is conducive to improving the resolution and reliability of geomagnetic depth sounding.

The magnetotelluric (MT) method is widely applied in petroleum, mining, and deep Earth structure exploration but suffers from cultural noise. This noise will distort apparent resistivity and phase, leading to false geological interpretation. Therefore, denoising is indispensable for MT signal processing. The sparse representation method acts as a critical role in MT denoising. However, this method depends on the sparse assumption leading to inadequate denoising results in some cases. We propose an alternative MT denoising approach, which can achieve accurate denoising without assumptions on datasets. We first design a residual network (ResNet), which has an excellent fitting ability owing to its deep architecture. In addition, the ResNet network contains skip-connection blocks to guarantee the robustness of network degradation. As for the number of training, validation, and test datasets, we use 10,000,000; 10,000; and 100 field data, respectively, and apply the gradual shrinkage learning rate to ensure the ResNet’s generalization. In the noise identification stage, we use a small-time window to scan the MT time series, after which the gramian angular field (GAF) is applied to help identify noise and divide the MT time series into noise-free and noise data. We keep the noise-free data section in the denoising stage, and the noise data section is fed into our network. In our experiments, we test the performances of different time window sizes for noise identification and suppression and record corresponding time consumption. Then, we compare our approach with sparse representation methods. Testing results show that our approach can obtain the desired denoising results. The accuracy and loss curves show that our approach can well suppress the MT noise, and our network has a good generalization. To further validate our approach’s effectiveness, we show the apparent resistivity, phase, and polarization direction of test datasets. Our approach can adjust the distortion of apparent resistivity and phase and randomize the polarization direction distribution. Although our approach requires the high quality of the training dataset, it achieves accurate MT denoising after training and can be meaningful in cases of a severe MT noisy environment.