Seismic waves from earthquakes and other sources are used to infer the structure and properties of Earth’s interior. The availability of large-scale seismic datasets and the suitability of deep-learning techniques for seismic data processing have pushed deep learning to the forefront of fundamental, long-standing research investigations in seismology. However, some aspects of applying deep learning to seismology are likely to prove instructive for the geosciences, and perhaps other research areas more broadly. Deep learning is a powerful approach, but there are subtleties and nuances in its application. We present a systematic overview of trends, challenges, and opportunities in applications of deep-learning methods in seismology. Description Large-scale learning The large amount and availability of datasets in seismology creates a great opportunity to apply machine learning and artificial intelligence to data processing. Mousavi and Beroza provide a comprehensive review of the deep-learning techniques being applied to seismic datasets, covering approaches, limitations, and opportunities. The trends in data processing and analysis can be instructive for geoscience and other research areas more broadly. —BG The ways in which deep learning can help process and analyze large seismological datasets are reviewed. BACKGROUND Seismology is the study of seismic waves to understand their origin—most obviously, sudden fault slip in earthquakes, but also explosions, volcanic eruptions, glaciers, landslides, ocean waves, vehicular traffic, aircraft, trains, wind, air guns, and thunderstorms, for example. Seismology uses those same waves to infer the structure and properties of planetary interiors. Because sources can generate waves at any time, seismic ground motion is recorded continuously, at typical sampling rates of 100 points per second, for three components of motion, and on arrays that can include thousands of sensors. Although seismology is clearly a data-rich science, it often is a data-driven science as well, with new phenomena and unexpected behavior discovered with regularity. And for at least some tasks, the careful and painstaking work of seismic analysts over decades and around the world has also made seismology a data label–rich science. This facet makes it fertile ground for deep learning, which has entered almost every subfield of seismology and outperforms classical approaches, often dramatically, for many seismological tasks. ADVANCES Seismic wave identification and onset-time, first-break determination for seismic P and S waves within continuous seismic data are foundational to seismology and are particularly well suited to deep learning because of the availability of massive, labeled datasets. It has received particularly close attention, and that has led, for example, to the development of deep learning–based earthquake catalogs that can feature more than an order of magnitude more events than are present in conventional catalogs. Deep learning has shown the ability to outperform classical approaches for other important seismological tasks as well, including the discrimination of earthquakes from explosions and other sources, separation of seismic signals from background noise, seismic image processing and interpretation, and Earth model inversion. OUTLOOK The development of increasingly cost-effective sensors and emerging ground-motion sensing technologies, such as fiber optic cable and accelerometers in smart devices, portend a continuing acceleration of seismological data volumes, so that deep learning is likely to become essential to seismology’s future. Deep learning’s nonlinear mapping ability, sequential data modeling, automatic feature extraction, dimensionality reduction, and reparameterization are all advantageous for processing high-dimensional seismic data, particularly because those data are noisy and, from the point of view of mathematical inference, incomplete. Deep learning for scientific discovery and direct extraction of insight into seismological processes is clearly just getting started. Aspects of seismology pose interesting additional challenges for deep learning. Many of the most important problems in earthquake seismology—such as earthquake forecasting, ground motion prediction, and rapid earthquake alerting—concern large and damaging earthquakes that are (fortunately) rare. That rarity poses a fundamental challenge for the data-hungry methods of deep learning: How can we train reliable models, and how do we validate them well enough to rely on them when data are scarce and opportunities to test models are infrequent? Further, how can we operationalize deep-learning techniques in such a situation, when the mechanisms by which they make predictions from data may not be easily explained, and the consequences of incorrect models are high? Incorporating domain knowledge through physics-based and explainable deep learning and setting up standard benchmarking and evaluation protocols will help ensure progress, as is the nascent emergence of a seismological data science ecosystem. More generally, a combination of data science literacy for geoscientists as well as recruiting data science expertise will help to ensure that deep-learning seismology reaches its full potential. Deep-learning processing of seismic data and incorporation of domain knowledge can lead to new capabilities and new insights across seismology. A. MASTIN/SCIENCE, TOP RIGHT: CARA HARWOOD/UNIVERSITY OF CALIFORNIA-DAVIS/CC BY-NC-SA 3.0 MIDDLE RIGHT: JOHAN SWANEPOEL/SCIENCE SOURCE

Deep-learning seismology


 The importance of low-frequency seismic data has been already recognized by geophysicists. However, there are still a number of obstacles that must be overcome for events recovery and noise suppression in low-frequency seismic data. The most difficult one is how to increase the signal-to-noise ratio (SNR) at low frequencies. Desert seismic data are a kind of typical low-frequency seismic data. In desert seismic data, the energy of low-frequency noise (including surface wave and random noise) is strong, which largely reduces the SNR of desert seismic data. Moreover, the low-frequency noise is non-stationary and non-Gaussian. In addition, compared with seismic data in other regions, the spectrum overlaps between effective signals and noise is more serious in desert seismic data. These all bring enormous difficulties to the denoising of desert seismic data and subsequent exploration work including geological structure interpretation and forecast of reservoir fluid. In order to solve this technological issue, feed-forward denoising convolutional neural networks (DnCNNs) are introduced into desert seismic data denoising. The local perception and weight sharing of DnCNNs make it very suitable for signal processing. However, this network is initially used to suppress Gaussian white noise in noisy image. For the sake of making DnCNNs suitable for desert seismic data denoising, comprehensive corrections including network parameter optimization and adaptive noise set construction are made to DnCNNs. On the one hand, through the optimization of denoising parameters, the most suitable network parameters (convolution kernel、patch size and network depth) for desert seismic denoising are selected; on the other hand, based on the judgement of high-order statistic, the low-frequency noise of processed desert seismic data is used to construct the adaptive noise set, so as to achieve the adaptive and automatic noise reduction. Several synthetic and actual data examples with different levels of noise demonstrate the effectiveness and robustness of the adaptive DnCNNs in suppressing low-frequency noise and preserving effective signals.

Desert low-frequency noise suppression by using adaptive DnCNNs based on the determination of high-order statistic

This review provides a comprehensive overview of the state-of-the-art methods of graph-based networks from a deep learning perspective. Graph networks provide a generalized form to exploit non-euclidean space data. A graph can be visualized as an aggregation of nodes and edges without having any order. Data-driven architecture tends to follow a fixed neural network trying to find the pattern in feature space. These strategies have successfully been applied to many applications for euclidean space data. Since graph data in a non-euclidean space does not follow any kind of order, these solutions can be applied to exploit the node relationships. Graph Neural Networks (GNNs) solve this problem by exploiting the relationships among graph data. Recent developments in computational hardware and optimization allow graph networks possible to learn the complex graph relationships. Graph networks are therefore being actively used to solve many problems including protein interface, classification, and learning representations of fingerprints. To encapsulate the importance of graph models, in this paper, we formulate a systematic categorization of GNN models according to their applications from theory to real-life problems and provide a direction of the future scope for the applications of graph models as well as highlight the limitations of existing graph networks.

/pdf/graph-neural-network-a-comprehensive-review-on-non-euclidean-3n5mtzw02g.pdf

Graph Neural Network: A Comprehensive Review on Non-Euclidean Space

Lots of low-frequency noise including random noise and surface waves seriously reduces the quality of desert seismic data. However, the suppression for desert low-frequency noise faces three main problems: nonstationary and non-Gaussian of random noise; strong energy of low-frequency noise; a more serious frequency-band overlap between effective signals and low-frequency noise. Robust principal component analysis (RPCA) is a classical low-rank matrix (LM) recovery method which is very suitable for processing nonlinear noise. It can decompose noisy data to the optimal LM and sparse matrix (SM), which include most effective signals and noise, respectively. Therefore, the RPCA is introduced to suppress desert low-frequency noise. However, due to the low signal-to-noise ratio (SNR) and serious frequency-band overlap, much low-frequency noise still remains in the LM of desert seismic data after the decomposition of RPCA. Meanwhile, some nonnegligible effective signals are decomposed into the SM of desert seismic data. To solve this problem, the convolutional neural network (CNN) is introduced to extract effective signals from SM and LM. By constructing suitable training sets to guide the CNN’s training, the CNN denoising models after training are used to predict the effective signals from these two matrices, respectively. In this article, to approach real desert seismic data, we use a variety of seismic wavelets to simulate different types of seismic events, and then use these synthetic seismic events and real desert low-frequency noise to construct training set. In experiments, our method can raise the SNR of synthetic noisy data from −8.69 to 9.63 dB.

New Suppression Technology for Low-Frequency Noise in Desert Region: The Improved Robust Principal Component Analysis Based on Prediction of Neural Network

Seismic exploration is a kind of exploration method for oil and gas resources. However, the disturbance of numerous random noise will decrease the quality and signal-to-noise ratio (SNR) of real seismic records, which brings difficulties to the following works of processing and interpretation. The seismic records of desert region pose a particular problem because of the strong energy noise and the spectrum overlapping between effective signals and random noise. Recent research works demonstrate that a convolutional neural network (CNN) can increase the SNR of seismic records. The optimization of denoising methods based on CNN is principally driven by the loss functions that largely focus on minimizing the mean-squared reconstruction error between denoising records and theoretical pure records. The denoising results estimated by the CNN model are often lacking the perfection of the signal structure. Therefore, when processing seismic records with low SNR, the denoising results often have a lack of effective signal in some traces, which leads to the poor continuity of events. In order to solve this problem, we adopt the strategy of generative adversarial network (GAN) to construct a GAN for denoising. It is divided into two parts: the generator (the denoising network based on CNN) is used to remove noise, while the discriminator is used to guide the generator to restore the structure information of effective signals. The generator and discriminator enhance the performance of each other through adversarial training, and the generator after adversarial training can greatly recover events and suppress random noise in synthetic and real desert seismic data.

Generative Adversarial Network for Desert Seismic Data Denoising

Distributed optical fiber acoustic sensing (DAS) is a new and rapid-developing detection technology in seismic exploration. Unfortunately, due to the weak energy of scattered optical signals and the inferior coupling between DAS cable and receiving interface, the seismic data received by DAS are often characterized by low signal-to-noise ratio (SNR); this low SNR is likely to affect some subsequent analysis, such as inversion, imaging, and interpretation. In addition, the noise caused by the inferior coupling is a new kind of noise not presented on conventional seismic data. To enhance the SNR of DAS seismic data and suppress the DAS noise effectively, we propose a convolutional adversarial denoising network (CADN) based on the basic strategy of generative adversarial network (GAN) and the usage of a denoiser to replace the original generator in GAN. In CADN, the performance of denoiser is significantly strengthened via its own mean square error (MSE) loss and the adversarial loss between it and the discriminator. To balance the two losses and thus ensure the optimization of denoiser, we construct a novel loss function, where the optimal ratio of MSE and adversarial losses is determined by quantifying the denoising performance. Both real and synthetic examples are included to testify the denoising performance of CADN. Experimental results have demonstrated that CADN can suppress most of the DAS noise and enhance the SNR of DAS seismic data; also, it can recover the effective signals completely, even the extremely weak effective signals reflected by deep layers.

Denoising the Optical Fiber Seismic Data by Using Convolutional Adversarial Network Based on Loss Balance

The seismic data with high quality are the essential foundation of imaging and interpretation. However, the real seismic data are inevitably contaminated by noise, which affects the subsequent processing and interpretation of seismic data. In desert seismic data, the energy of noise is stronger. Also, the frequency-band overlap between noise and effective signals is more serious. Recently, some methods based on supervised learning can suppress the desert seismic noise to some extent. Generally, supervised learning-based methods use synthetic noisy data and paired pure data as training sets to train model. However, the difference between synthetic noisy data of training and real seismic data of testing leads to the degradation of the model, and the denoising results often have many false seismic events when dealing with field seismic data. To solve the above problem, we introduce Cycle-generative adversarial networks (GANs) into the denoising of desert seismic records. Cycle-GAN is an unsupervised learning-based method. It can learn the domain mapping from noisy data domain to effective signal data domain through unpaired data training. So we use unpaired real desert common-shot-point data and synthetic pure data to train Cycle-GAN, so as to effectively improve the denoising ability of the method for real seismic data. Finally, the denoising of desert seismic data is realized. The experiment shows that the Cycle-GAN with unpaired data training can effectively suppress desert seismic noise and retain the effective signal amplitude. Also, the denoising result has less false seismic reflection.

The Denoising of Desert Seismic Data Based on Cycle-GAN With Unpaired Data Training

Seismic prospecting is a common method used in oil and gas resource exploration. However, due to the limitations of current collection techniques, seismic records acquired in the field are typically contaminated by severe incoherent noise, which has negative implications for the subsequent processing and interpretation procedures. In addition, numerous traditional denoising algorithms have been applied in order to mitigate this problem, but further improvements are required, especially for the seismic data with spectral overlapping between effective signals and background noise. In recent years, feedforward denoising convolutional neural networks (DnCNNs) have been applied to suppress the complex random noise, and a series of essential insights have been gained. Nonetheless, conventional denoising networks always extract data features depending on single-scale information, resulting in impaired performance when coping with seismic records with a low signal-to-noise ratio (SNR). For solving this problem, a novel multiscale DnCNN (MSDCNN) is developed as an attempt for random noise suppression. Unlike conventional DnCNN, MSDCNN has a hierarchical structure capable of extracting features at different scales and capturing informative and discriminatory features through effective information integration. Meanwhile, the cross-scale feature interaction also increases the processing accuracy when confronted with weak reflection events. Experimental results derived from both synthetic and field data indicate that the proposed network can effectively suppress the random noise and accurately preserve reflection events, even under low SNR conditions.

Xintong Dong

Papers

Denoising the Optical Fiber Seismic Data by Using Convolutional Adversarial Network Based on Loss Balance

New Suppression Technology for Low-Frequency Noise in Desert Region: The Improved Robust Principal Component Analysis Based on Prediction of Neural Network

Generative Adversarial Network for Desert Seismic Data Denoising

The Denoising of Desert Seismic Data Based on Cycle-GAN With Unpaired Data Training

Seismic Random Noise Attenuation by Applying Multiscale Denoising Convolutional Neural Network