The magnetotelluric (MT) data collected in an ore-concentration area are extremely vulnerable to all kinds of noise pollution. However, separating real MT signals from strong noise is still a difficult problem, and the noise in MT data is quite distinct from clean data in morphological features. By performing the signal-noise identification and data prediction, we develop a deep learning method to denoise MT data containing strong noise. First, we use the convolutional neural network (CNN) to learn the feature differences between the samples of massive noise and clean data and use the learned features to realize signal-noise identification of the measured data. Second, we use the measured clean data obtained by CNN identification to train the long short-term memory (LSTM) neural network and perform the prediction denoising of the noisy data. The simulation results clearly demonstrate the following two facts: (1) the predicted data output from LSTM basically matches the time-frequency domain features of the real data and (2) our CNN method performs significantly better than the features parameter classification method in dealing with signal-noise identification. In addition, the validity of our method is verified by the processing results of the measured data.

It is well known that a magnetotelluric (MT) signal with high signal-to-noise ratio is an important prerequisite for correct interpretation of subsurface structures. However, MT signals collected in the environment of strong cultural noise often are of low data quality due to noise pollution, which seriously affects the accuracy of interpretation. As can be seen from the MT time-domain waveform, the noise is highly energetic, diverse, and random. This means MT denoising methods should have strong applicability to guarantee accurate and effective separation of MT signal from noise data. Therefore, we propose a deep-learning-based data nonlinear mapping method for MT signal-to-noise separation. First, this method focuses on learning the nonlinear mapping relationship between a large amount of noise data and the corresponding noise contour by using the convolutional neural network (CNN) in advance. Then, the mapping transformation of noise data to noise contour in the measured data is realized by CNN model. Specifically, based on the features of MT noise data, we construct a large amount of training data very close to it by mathematical functions. At the same time, we also select some of the measured data to be added to the training set. This not only expands the amount and diversity of the training set but also improves the adaptability of CNN when dealing with complex data. Finally, we evaluate the denoising performance of the proposed method in terms of time-domain waveforms, apparent resistivity-phase curves, and polarization directions before and after denoising. The processing results of the simulated data and the measured data collected in Luzong area have verified the feasibility and effectiveness of the proposed method in MT data denoising.

Audio magnetotellurics (AMT), as a commonly used passive geophysical technique, provides outstanding metal ore exploration capabilities based on the resistivity structure of the earth. However, the accuracy of AMT in translating geoelectrical structures decreases when the data collected in mining areas are of poor quality and contain complex anthropogenic noise, leading to distorted apparent resistivity-phase curves and posing significant challenges for mineral exploration. To effectively denoise AMT data, we develop a new denoising method that combines atom-profile updating dictionary learning (APrU) with the nucleus sampling attention mechanism (NSAM) sparse coding. First, we use APrU to accurately learn the characteristics of the noise in the AMT data; then, we apply the updated dictionary to perform sparse coding on the AMT data by NSAM to obtain the noise; finally, we subtract the noise from the original AMT data to obtain the denoised data. Our experimental results suggest that the proposed method can learn an overcomplete dictionary via the to-be-processed AMT data, thereby enabling the sparse representation of the noise within the learned dictionary. We also demonstrate the efficacy of this method with a set of field data collected from the Lu-Zong mining area, and the attained denoised data faithfully restore the geoelectrical structures with heightened accuracy. The findings confirm that the proposed method realizes the unsupervised learning of the AMT data and allows us to achieve precise denoising performance.

Strong noise is one of the biggest challenges in controlled-source electromagnetic (CSEM) exploration, which severely affects the quality of the recorded signal. We develop a novel and effective CSEM noise attenuation method that integrates a deep convolutional neural network (DCNN) with a high-quality sounding curve screening mechanism. This method first uses a DCNN to learn the large-scale noise characteristics of CSEM measured data and build a DCNN denoising model for attenuating the noise. Because some impulse noise remains after DCNN denoising, the high-quality sounding curve screening mechanism is innovatively applied to optimize the DCNN denoised data. The core of this method is the smoothness value (SV), which quantitatively evaluates the smoothness of the sounding curve (apparent resistivity curve). In effect, the mechanism segments the overall DCNN denoised data and calculates the corresponding sounding curve and its SV for each segment. The highest-quality sounding curve can be screened out by comparing the SVs of all curves as well as obtaining the highest-quality data segment present in the overall DCNN denoised data. Finally, our method is validated with synthetic and field data, demonstrating its feasibility and effectiveness in attenuating large-scale CSEM noise. The quality of CSEM data, which is heavily affected by noise, is significantly improved after processing by our method.

Seismic monitoring has been instrumental in various domains such as natural earthquake early warning, mineral mining safety assessment, and hydraulic fracturing impact evaluation. However, the monitoring data often exhibit low signal-to-noise ratio (SNR) and large volume. Therefore, developing an efficient, high-precision, and universally applicable seismic waveform automatic classification network model becomes significant and practical. We propose a physical interpretable time-frequency deep convolutional recurrent neural network (TF-DCRNN) model which consists of an integration of a time-frequency convolutional (TFconv) layer and a convolutional recurrent neural network (CRNN). Subsequently, we evaluate the classification performance by comparing five network models, including convolutional neural network (CNN) and long short-term memory (LSTM), using Ricker wavelet datasets with varying SNR levels (-15 ~ 0 dB). Our findings verify the superiority of the TF-DCRNN model in the classification of strong interference environment from both numerical and physical simulation. Moreover, integrating multiple network models or incorporating a TFconv layer can moderately enhance the classification performance, which provides the direction for network model optimization.