Recently deep learning (DL), as a new data-driven technique compared to conventional approaches, has attracted increasing attention in geophysical community, resulting in many opportunities and challenges. DL was proven to have the potential to predict complex system states accurately and relieve the “curse of dimensionality” in large temporal and spatial geophysical applications. We address the basic concepts, state-of-the-art literature, and future trends by reviewing DL approaches in various geosciences scenarios. Exploration geophysics, earthquakes, and remote sensing are the main focuses. More applications, including Earth structure, water resources, atmospheric science, and space science, are also reviewed. Additionally, the difficulties of applying DL in the geophysical community are discussed. The trends of DL in geophysics in recent years are analyzed. Several promising directions are provided for future research involving DL in geophysics, such as unsupervised learning, transfer learning, multimodal DL, federated learning, uncertainty estimation, and active learning. A coding tutorial and a summary of tips for rapidly exploring DL are presented for beginners and interested readers of geophysics.

https://agupubs.onlinelibrary.wiley.com/doi/pdf/10.1029/2021RG000742

Deep learning for geophysics: Current and future trends

Seismic reflection data were used to monitor the progress of an in‐situ combustion, enhanced oil recovery process. Three sets of three‐dimensional (3-D) data were collected during a one‐year period in order to map the extent and directions of propagation in time. Acquisition and processing parameters were identical for each survey so that direct one‐to‐one comparison of traces could be made. Seismic attributes were calculated for each common‐depth‐point data set, and in a unique application of seismic reflection data, the preburn attributes were subtracted from the midburn and postburn attributes. The resulting “difference volumes” of 3-D seismic data showed anomalies which were the basis for the interpretation shown in this case study. Profiles and horizon slices from the data sets clearly show the initiation and development of a bright spot in the reflection from the top of the reservoir and a dim spot in the reflection from a limestone below it. Interpretation of these anomalies is supported by informa...

AAPG Memoir 42 and SEG Investigations in Geophysics, No. 9, Chapter 9 (Case Histories of Three-Dimensional Seismic Surveys) -- Case History 4: Three-Dimensional Seismic Monitoring of An Enhanced Oil Recovery Process

Template matching is an alternative to perform speech recognition. However, the template matching encountered problems due to speaking rate variability, in which there exist timing differences between the two utterances. Speech has a constantly changing signal, thus it is almost impossible to get the same signal for two same utterances. The problem of time differences can be solved through DTW algorithm: warping the template with the test utterance based on their similarities. So, DTW algorithm actually is a procedure, which combines both warping and distance measurement. DTW is considered as one effective method in speech pattern recognition, however the bad side of this method is that it requires a long processing time plus large storage capacity, especially for real time recognitions.

Dynamic time warping

This study presents the first demonstration of the transferability of a convolutional neural network (CNN) trained to detect microseismic events in one fiber-optic distributed acoustic sens...

/pdf/application-of-machine-learning-to-microseismic-event-32ugvz8rlg.pdf

Application of machine learning to microseismic event detection in distributed acoustic sensing data

An immediate report of the source focal mechanism with full automation after a destructive earthquake is crucial for timely characterizing the faulting geometry, evaluating the stress perturbation, and assessing the aftershock patterns. Advanced technologies such as Artificial Intelligence (AI) has been introduced to solve various problems in real-time seismology, but the real-time source focal mechanism is still a challenge. Here we propose a novel deep learning method namely Focal Mechanism Network (FMNet) to address this problem. The FMNet trained with 787,320 synthetic samples successfully estimates the focal mechanisms of four 2019 Ridgecrest earthquakes with magnitude larger than Mw 5.4. The network learns the global waveform characteristics from theoretical data, thereby allowing the extensive applications of the proposed method to regions of potential seismic hazards with or without historical earthquake data. After receiving data, the network takes less than two hundred milliseconds for predicting the source focal mechanism reliably on a single CPU. The authors here present a deep learning method to determine the source focal mechanism of earthquakes in realtime. They trained their network with approximately 800k synthetic samples and managed to successfully estimate the focal mechanism of four 2019 Ridgecrest earthquakes with magnitudes larger than Mw 5.4.

/pdf/real-time-determination-of-earthquake-focal-mechanism-via-3t46qlets4.pdf

Real-time determination of earthquake focal mechanism via deep learning.

The accurate and automated determination of earthquake locations is still a challenging endeavor. However, such information is critical for monitoring seismic activity and assessing potential hazards in real time. Recently, a convolutional neural network was applied to detect earthquakes from single-station waveforms and approximately map events across several large surface areas. In this study, we locate 194 earthquakes induced during oil and gas operations in Oklahoma, USA, within an error range of approximately 4.9 km on average to the epicenter and 1.0 km to the depth in catalogs with data from 30 network stations by applying the fully convolutional network. The network is trained by 1,013 historic events, and the output is a 3D volume of the event location probability in the Earth. The trained system requires approximately one hundredth of a second to locate an event without the need for any velocity model or human interference.

/pdf/locating-earthquakes-with-a-network-of-seismic-stations-via-2zy32nd9um.pdf

Locating earthquakes with a network of seismic stations via a deep learning method.

The accurate and automated determination of small earthquake (ML   ML ≥ 0.5), and the output is a 3D volume of the event location probability in the Earth. The prediction results suggest that the mean epicenter errors of the testing events (ML ≥ 1.5) vary from 3.7 to 6.4 km, meeting the need of the traffic light system in Oklahoma, but smaller events (ML = 1.0, 0.5) show errors larger than 11 km. Synthetic tests suggest that the accuracy of ground truth from catalog affects the prediction results. Correct ground truth leads to a mean epicenter error of 2.0 km in predictions, but adding a mean location error of 6.3 km to ground truth causes a mean epicenter error of 4.9 km. The automated system is able to distinguish certain interfered events or events out of the monitoring zone based on the output probability estimate. It requires approximately one hundredth of a second to locate an event without the need for any velocity model or human interference.

/pdf/locating-induced-earthquakes-with-a-network-of-seismic-59tl9dbt7r.pdf

Locating induced earthquakes with a network of seismic stations in Oklahoma via a deep learning method.

\n Temporal changes in subsurface properties, such as seismic wave speeds, can be monitored by measuring phase shifts in the coda of two seismic waveforms that share a similar source–receiver path but that are recorded at different times. These nearly identical seismic waveforms are usually obtained either from repeated earthquake waveforms or from repeated ambient noise cross-correlations. The five algorithms that are the most popular to measure phase shifts in the coda waves are the windowed cross correlation (WCC), trace stretching (TS), dynamic time warping (DTW), moving window cross spectrum (MWCS) and wavelet cross spectrum (WCS). The seismic wave speed perturbation is then obtained from the linear regression of phase shifts with their respective lag times under the assumption that the velocity perturbation is homogeneous between (virtual or active) source and receiver. We categorize these methods into the time domain (WCC, TS, DTW), frequency domain (MWCS) and wavelet domain (WCS). This study complements this suite of algorithms with two additional wavelet-domain methods, which we call wavelet transform stretching (WTS) and wavelet transform DTW, wherein we apply traditional stretching and DTW techniques to the wavelet transform. This work aims to verify, validate, and test the accuracy and performance of all methods by performing numerical experiments, in which the elastic wavefields are solved for in various 2-D heterogeneous half-space geometries. Through this work, we validate the assumption of a linear increase in phase shifts with respect to phase lags as a valid argument for fully homogeneous and laterally homogeneous velocity changes. Additionally, we investigate the sensitivity of coda waves at various seismic frequencies to the depth of the velocity perturbation. Overall, we conclude that seismic wavefields generated and recorded at the surface lose sensitivity rapidly with increasing depth of the velocity change for all source–receiver offsets. However, measurements made over a spectrum of seismic frequencies exhibit a pattern such that wavelet methods, and especially WTS, provide useful information to infer the depth of the velocity changes.

Numerical comparison of time-, frequency- and wavelet-domain methods for coda wave interferometry


 It is of great significance and a great challenge to quickly and effectively monitor subsurface time-lapse velocities in the earth. Over the past few decades, regularized iterative methods, such as traveltime and waveform inversions, have been presented to monitor velocity changes. Due to high processing cost, these iterative methods have been hardly employed in practice to investigate the subsurface velocity changes in real time. In this study, we propose a new time-lapse imaging technique that effectively eliminates these limitations and directly produces accurate velocity changes from the time-lapse data. The approach uses a fully convolutional neural network (FCN) to perform the inverse problem. The network architecture consists of a contracting path to quickly extract the features of waveform data and a symmetric expanding path to yield an accurate velocity model. With the known baseline velocity and data, we cast a mapping between time-lapse data and target velocity changes via the proposed FCN algorithm. Along with the observed time-lapse data, this mapping will generate a predictive estimation of the target velocity changes. We demonstrate the efficiency and accuracy of our approach in three 2D synthetic tests. The proposed technique is able to invert the velocity changes successfully with much higher efficiency than the regular double-difference full waveform inversion.

Congcong Yuan

Papers

Locating earthquakes with a network of seismic stations via a deep learning method.

Real-time determination of earthquake focal mechanism via deep learning.

Locating induced earthquakes with a network of seismic stations in Oklahoma via a deep learning method.

Numerical comparison of time-, frequency- and wavelet-domain methods for coda wave interferometry

Time-lapse velocity imaging via deep learning