Field Data Application Research Articles

Deep learning-based point-spread function deconvolution for migration image deblurring

High-resolution seismic imaging is crucial for deriving an accurate geologic section and ensuring successful petroleum exploration. However, traditional high-resolution imaging methods, such as least-squares reverse time migration (LSRTM) and migration deconvolution, usually involve high computational costs for approximating the inverse Hessian matrix with limited improvements in resolution. To obtain high-resolution migration images and reduce computing costs, we present a deep learning-based point-spread function (PSF) deconvolution method. We decompose the large Hessian matrix into small dispersed PSFs and design a convolutional neural network (CNN) to automatically predict the deconvolution operator of every single PSF. The deconvolution operator eliminates the PSF effect and balances the insufficient illumination of the acquisition geometry. To obtain an efficient CNN model for predicting deconvolution operators, we trained 2500 pairs of PSFs and their corresponding deconvolution operators collected from part of a model using the regular PSF deconvolution method. After training the network, we could predict all deconvolution operators in a few seconds for the following processing of migration images. The results on synthetic and field-data applications indicate that our method could provide a deblurred migration image comparable to that of the LSRTM approaches with significantly reduced computational and memory costs.

Dynamic subwindow matching: A new similarity measure for seismic facies analysis

Dynamic subwindow matching: A new similarity measure for seismic facies analysis

Elastic Transmitted Wave Reverse Time Migration for Imaging Earth’s Interior Discontinuities: A Numerical Study

ABSTRACT Seismic imaging is crucial in investigating Earth’s interior structures and understanding its tectonics and evolution. The reflected, converted, and scattered waves have attracted considerable attention in the previous studies, whereas the directly transmitted waves are less used in seismic imaging. In this study, we present a novel passive source elastic transmitted wave reverse time migration (T-RTM) method to characterize major discontinuities in Earth’s interior using transmitted P or S waveforms. By extrapolating and then cross correlating the wavefields from the sources with the transmitted wavefields from the receivers using flood velocity models, the velocity discontinuities can be clearly imaged. The advantages and potential applications of the proposed T-RTM method are demonstrated with three synthetic imaging experiments. First, with distributed acoustic sensing using submarine cables, we show that the proposed method can be used to image the Moho with teleseismic S waves recorded in a single axial component, which is difficult for other imaging methods. Second, using direct P waveforms with a single P-wave velocity model, we show that the proposed method can be used to image 3D Moho topography without relying on the VP/VS ratio like in the traditional receiver function imaging. Third, the proposed method can also be used to image a fault plane with a sharp velocity contrast using local earthquakes. We discuss the limitations of the proposed method and some potential issues in field-data applications with additional numerical experiments. The proposed T-RTM method could provide many new opportunities for utilizing transmitted waveforms in the study of oceanic and continental structures.

Iterative deblending using MultiResUNet with multilevel blending noise for training and transfer learning

Blended seismic acquisition has improved the efficiency of land and marine data acquisition significantly. Nevertheless, the consequent blending noise poses challenges for subsequent seismic imaging and inversion. Therefore, deblending algorithms are being widely investigated. To improve the deblending performance and efficiency of traditional deblending algorithms, a new method is proposed that we call multiresolution ResUNet (MultiResUNet) trained on data sets with multilevel blending noise. The trained MultiResUNet is optimal for iterative deblending. MultiResUNet combines the advantages of a residual learning network (ResNet) and U-net. We apply a scalar factor to the synthetic blending noise to construct a training data set with multilevel blending noise contamination, which is designed to train and fine-tune MultiResUNet to detect weak blending noise and effectively retrieve signals in an iterative manner. The proposed method attenuates the blending noise iteratively, leading to an improved performance compared with the conventional curvelet transform-based thresholding algorithm or MultiResUNet trained on data sets with single-level blending noise. Because the training and fine tuning of MultiResUNet happens only once up front, the application of the trained MultiResUNet is efficient. To demonstrate the performance of the proposed method, the improved deblending accuracy is verified through comparison on numerical examples. For field data applications, a transfer learning approach is adopted to generalize the MultiResUNet trained on synthetic blended data for accurate deblending.

Robust full-waveform inversion with graph-space optimal transport: Application to 3D ocean-bottom cable Valhall data

Improving full-waveform inversion robustness to cycle-skipping has been the subject of a large number of studies. From the several families of approaches developed, one of the most documented consists in modifying the least-squares distance misfit function. From all the propositions made to improve and replace the least-squares distance, only a few of them have been applied to field data. One of the methods proposed recently, the graph space optimal transport distance, presents appealing properties for field data applications. We compare it with the least-squares distance in an analysis performed on the three-dimensional ocean bottom cable data from the Valhall field. We first perform this comparison starting the inversion from the reflection traveltime tomography model used in previous studies. We then perform a second comparison from a crude, linearly varying in-depth one-dimensional velocity model. Starting from this model, least-squares-based full-waveform inversion fails to provide a meaningful estimate of the pressure-wave velocity model due to cycle skipping. We illustrate how the graph-space optimal transport-based full-waveform inversion mitigates this issue. A meaningful estimate of the pressure-wave velocity model is obtained in the zone sampled by both diving and reflected waves, down to almost two kilometers depth.

Deep Learning for Efficient Microseismic Location Using Source Migration‐Based Imaging

AbstractMigration‐based location methods (e.g., time‐reverse imaging based on wave equation, Kirchhoff summation, and diffraction stacking) can effectively locate events of low signal‐to‐noise ratios by stacking waveforms from many receivers. The methods have been widely applied for surface microseismic monitoring. However, these methods may not produce accurate results if there are polarity reversals in the surface records for a double‐couple or even a general moment tensor event. Various imaging conditions have been developed to solve the non‐focus image problem for a non‐explosive source. Here, we propose a deep convolutional neural network to predict a better‐focused image from a regular migration image that contains a quasi‐symmetric pattern in both space and time. To train the network, we first simulate a large number of surface records from sources with various locations and mechanisms. We then compute diffraction stacking images from the records and take the images as the input to the network. We define the corresponding training labels as images (with the same size as the input) with Gaussian distributions centered at the true sources. This network, trained by only synthetic datasets, works well in field data to detect source locations from images for unknown events. Both synthetic tests and field data applications demonstrate that the proposed method can effectively improve diffraction stacking images for efficient microseismic location.

Elastic least-squares reverse time migration of steeply dipping structures using prismatic reflections

The migration of prismatic reflections can be used to delineate steeply dipping structures, which is crucial for oil and gas exploration and production. Elastic least-squares reverse time migration (ELSRTM), which considers the effects of elastic wave propagation, can be used to obtain reasonable subsurface reflectivity estimations and interpret multicomponent seismic data. In most cases, we can only obtain a smooth migration model. Thus, conventional ELSRTM, which is based on the first-order Born approximation, considers only primary reflections and cannot resolve steeply dipping structures. To address this issue, we develop an ELSRTM framework, called Pris-ELSRTM, which can jointly image primary and prismatic reflections in multicomponent seismic data. When Pris-ELSRTM is directly applied to multicomponent records, near-vertical structures can be resolved. However, the application of imaging conditions established for prismatic reflections to primary reflections destabilizes the process and leads to severe contamination of the results. Therefore, we further improve the Pris-ELSRTM framework by separating prismatic reflections from recorded multicomponent data. By removing artificial imaging conditions from the normal equation, primary and prismatic reflections can be imaged based on unique imaging conditions. The results of synthetic tests and field data applications demonstrate that the improved Pris-ELSRTM framework produces high-quality images of steeply dipping P- and S-wave velocity structures. However, it is difficult to delineate steep density structures because of the insensitivity of the density to prismatic reflections.

Absorption Attenuation Compensation Using an End-to-End Deep Neural Network

Absorption attenuation compensation is an important part of seismic data processing. It enhances the resolution of non-stationary seismic data by compensating the amplitude energy and correcting phase distortion. The stabilized inverse <i>Q</i> -filter method, a widely used attenuation compensation method, constructs compensation operators based on stratigraphy-related assumptions and compensates seismic data using time-window analysis, which is computationally complex and sensitive to noise. The essence of attenuation compensation lies in the establishment of a nonlinear mapping relationship between attenuated and non-attenuated seismic traces, which strongly benefits from deep learning. This paper proposes a new method for attenuation compensation based on an end-to-end deep neural network to reduce the hand-crafted step of time-window analysis. Instead, the convolutional blocks of the network automatically learn and process seismic data features to achieve simultaneous amplitude and phase compensation. We have constructed two end-to-end network architectures for attenuation compensation: a fully convolutional network (FCN) and a U-Net. As an effective spectrum-broadening method, the compensation method based on the U-Net is shown to enhance vertical resolution with good lateral continuity, to provide reliable compensation results without complex calculations, and to exhibit high noise robustness. Synthetic data tests indicate that the compensation results from the U-Net are better than those from either the FCN or the stabilized inverse <i>Q</i> -filter method at different noise levels. Moreover, the correlation coefficient between the U-Net compensation results of the synthetic profile and the reference non-attenuated profile is higher than that of the FCN and the stabilized inverse <i>Q</i> -filter method. A field data application further verifies the effectiveness of this method.

Separating P- and S-Waves Based on the Slope of Wavefield Events and Polarizability

In multicomponent surveys, the travel path of seismic waves from the subface interfaces to geophones may not be vertical to the ground; consequently, vertical and horizontal geophones receive mixed wavefields comprising P- and S-waves. Vector separation of wavefields is crucial for subsequent imaging and shear wave splitting analyses. However, conventional methods cannot address the nonlinear polarization of pure wave and can also produce incorrect results due to the presence of dipping interface. To address the above issues, we propose a vector wavefield separation method based on the slope of events and polarizability. The proposed method calculates the slope of the events for determining the propagation direction of the wavefield. The polarizability is used to evaluate if the resulting wavefield is purely composed of one type of body wave while the wavefield type is determined based on the angle between the polarization and propagation directions. When the wavefield is mixed, P/S separation is realized based on vector decomposition in an affine coordinate system optimized by propagation direction. We use both numerical and field data to demonstrate the effectiveness and applicability of the proposed method. The numerical example confirms that the proposed method yields more accurate results than the conventional vector decomposition method. The field data application also indicates that the new method can effectively resolve the P- and S-waves from the noisy seismic wavefield.

Hierarchical Bayesian Probabilistic Seismic AVO Inversion Using Gibbs Sampling With IA2RMS Algorithm

From the observed seismic data, the estimation of elastic parameters and lithofacies of subsurface layers with seismic amplitude variation with offset (AVO) inversion plays an important role in reservoir characterization. The sequential estimation of elastic parameters and lithofacies in AVO inversion usually affect the prediction accuracies of model parameters without considering the prior information related to lithofacies. We propose a novel probabilistic AVO inversion approach in hierarchical Bayesian framework that combines a hierarchical iterative strategy with Gibbs-IA2RMS sampling, to estimate elastic parameters and lithofacies simultaneously. From the exact PP-wave reflection coefficient, we first derive the posterior distribution of model parameters involving a Bayesian hyper-parameter of lithofacies through the hierarchical prior probability distributions and the likelihood distribution of observed data. Based on Monte Carlo Markov chain approximation, we then introduce an iterative Gibbs-IA2RMS sampling to achieve the high sampling stability of the high-dimensional posterior PDF in AVO inversion. The algorithm realizes the adaptive construction of proposal distribution, and has a higher acceptance rate than the generic Metropolis-Hastings algorithm. The numerical results show that the approach effectively simulates the posterior distributions and inversion uncertainties of model parameters, and improves the computational efficiency of probabilistic AVO inversion. The applicability and validity of the proposed approach are also demonstrated with a field data application.

Multiscale Coherence Attribute and Its Application on Seismic Discontinuity Description

Geologic structure characterization is a key step for seismic structure interpretation, such as fluvial channels, faults, and fractures. The coherence attribute is a widely used tool for describing seismic discontinuities, which is usually calculated based on the similarity and dissimilarity of the adjacent seismic traces. However, accurately extracting coherence attribute is a difficult task in field data applications because seismic signal is one of the typical nonstationary, non-Gaussian, and wideband signals. To describe seismic discontinuities at different scales, we propose a workflow to extract the multiscale coherence (MSC) attribute. We first decompose seismic data into several band-limited intrinsic mode functions (IMFs) with different dominant frequencies via the multichannel variational mode decomposition (MVMD). Afterward, we develop a Cauchy kernel correlation-based coherence algorithm to extract the coherence attributes at different scales based on the decomposed IMFs. Finally, we can compute the MSC attribute by utilizing the calculated coherence attributes. Field data applications demonstrate that the proposed MSC attribute characterizes seismic discontinuities, such as faults and fluvial channels, more accurately and more clearly than the traditional coherence attribute and the 1-D variational mode decomposition (VMD)-based coherence attribute.

Estimation of Seismic Quality Factor via Quantum Mechanics-Based Signal Representation

We propose a stable seismic Q estimation approach by employing quantum mechanics-based signal representation. For Q estimation, we project a seismic trace onto a specific basis composed of wave functions constructed through the resolution of the Schroedinger equation of non-relativistic quantum mechanics at first. Then, based on the specific basis, we derive the quantum mechanics-based Q estimation approach in the local frequency-projection coefficient domain. The Planck constant and the control factor are the two key factors for the quantum mechanics-based Q estimation method. Compared with the traditional methods, the quantum mechanics-based Q estimation approach shows more stability and noise robustness. The synthetic and field data applications illustrate the effectiveness and the superiority of the proposed method. The quantum mechanics-based Q estimation method offers a new field and provides a complementary way for measuring seismic attenuation.

Multi-Task Learning for Low-Frequency Extrapolation and Elastic Model Building From Seismic Data

Low-frequency signal content in seismic data as well as a realistic initial model are key ingredients for robust and efficient full-waveform inversions. However, acquiring low-frequency data is challenging in practice for active seismic surveys. Data-driven solutions show promise to extrapolate low-frequency data given a high-frequency counterpart. While being established for synthetic acoustic examples, the application of bandwidth extrapolation to field datasets remains non-trivial. Rather than aiming to reach superior accuracy in bandwidth extrapolation, we propose to jointly reconstruct low-frequency data and a smooth background subsurface model within a multi-task deep learning framework. We automatically balance data, model and trace-wise correlation loss terms in the objective functional and show that this approach improves the extrapolation capability of the network. We also design a pipeline for generating synthetic data suitable for field data applications. Finally, we apply the same trained network to synthetic and real marine streamer datasets and run an elastic full-waveform inversion from the extrapolated dataset.

Seismic Data Denoising With Correlation Feature Optimization Via S-Mean

Random noise elimination acts as an important role in seismic data processing. Moreover, protecting and recovering useful subsurface structure information is also significant. In this study, the S-mean which can obtain the geometric mean of the seismic traces on the symmetric positive definite matrix (SPD) manifold is adopted as a nonlinear filter for seismic denoising. Furthermore, S-mean has the best correlation with other elements based on the S-divergence due to the optimization of finding S-mean on SPD manifold. Therefore, the broken correlation features in noisy seismic data are compensated and maintained well, which can be conducive to describe the subsurface structures. Synthetic example and field data applications qualitatively and quantitatively demonstrate the validity and effectiveness of the proposed workflow.

Inversion of Fracture Weaknesses Based on Linearly Approximated Traveltime

Seismic traveltime is crucial for the inversion of anisotropy parameters. In the inversion for fractured rock, a typical transversely isotropic medium with a horizontal symmetry axis (HTI), the joint utilization of traveltime of P- and split shear waves can provide more anisotropy information than the utilization of only P-wave traveltime. The computational effort of nonlinear parameter inversion also increases with the involvement of split shear waves. Therefore, based on the series expansion of vertical slowness projection, we derive the linear approximations for the first-break traveltimes of P-, S1-, and S2-waves with respect to fracture weaknesses, which are used to accelerate and stabilize the inversion process. Linear coefficients are used to analyze the sensitivity of the traveltime to fracture weakness and thus the accuracy of the results. To make our proposed approximation more practical, we expand our linear equations to the traveltime difference between the HTI model and its isotropic background model and derive the joint inversion equation for fracture weaknesses in a vertical seismic profile (VSP) survey by jointly using P-, S1-, and S2-waves. Field data application shows that our method is feasible for estimating fracture weaknesses by using walkaround VSP data. The error caused by linear approximation can be reduced by the joint utilization of multiple wave modes.

EisNet: Extracting Bedrock and Internal Layers From Radiostratigraphy of Ice Sheets With Machine Learning

The ice–bedrock interfaces at the bottoms of ice sheets and the internal ice layers record the historical evolution of the ice sheets and are important indicators for inferring glacier dynamics and explaining subglacial topographies. Radar-detected bedrock and internal ice layers can be used to observe conditions in the interiors and bottoms of ice sheets. The low contrast of the internal layers’ feature in radar images, as well as the effects of ice flow, makes it challenging to automatically extract and trace the layers and interfaces of ice sheets. Manual or semiautomatic methods are extensively applied in bedrock and internal ice layers’ extractions. However, the conventional methods require large amounts of time. Especially when processing a large number of radar observation data, the accuracy and efficiency of conventional methods are untoward. Therefore, we propose EisNet, a fusion system consisting of deep neural networks, to extract the multiple types of internal layers. Because of the rareness of high-precision extraction of ice layers, it is virtually impossible to train EisNet in observational data with the corresponded label of ice layers. To train EisNet, we design a novel synthetic method for radar images based on artifact noise and characteristic textures that are similar to those in the observation data. The validation in synthetic data proves EisNet’s capacity for layer extraction and noise removal. The field data application shows the generalization and adaptation abilities of EisNet and indicates that EisNet can significantly improve the efficiency of the automatic processing of ice sheet radiostratigraphy.

Seismic Sparse Time-Frequency Network With Transfer Learning

Time-frequency analysis (TFA) is a powerful tool for describing time-frequency (TF) features of seismic data, such as short-time Fourier transform and S-transform. Recently, sparse time-frequency analysis (STFA) is proposed for enhancing TF readability of commonly used TFA tools. However, STFA is often solved via an optimal inverse problem with a prior regularization term, which is difficult to set in practice, where the key regularization parameters are sensitive to noise. Moreover, it often takes expensive calculation time, especially for 3D field data application. We build a deep learning based workflow for implementing STFA to obtain sparse time-frequency (STF) spectra, termed the sparse time-frequency network with transfer learning (STFNTL). We first adopt a Marmousi II reflectivity model and Ricker wavelets with different dominant frequencies to generate synthetic training data set. Then, we adopt a simplified STFA method with optimized parameters to generate synthetic training labels, i.e., sparse TF spectra. Afterward, we propose the sparse time-frequency network (STFN) based on a simplified Unet model, which is trained using synthetic training data and corresponding STF labels. Moreover, to enhance the generalization of STFN, we introduce an adaptive transfer learning strategy based on small samples of field data and their corresponding STF labels. Finally, synthetic and field data are utilized to illustrate the effectiveness and generalization ability of our proposed model.

Broad Learning Framework for Search Space Design in Rayleigh Wave Inversion

Search space design is a fundamental procedure for the inverse problem in Rayleigh wave exploration. In practice, such a crucial task mainly depends on researchers’ experience-based judgment. However, intricate near-surface materials may lead to erroneous search space design, and consequently, accurate inversion results cannot always be ensured. By the forward calculation, it is found that there is a strong relationship between the fundamental dispersion curves of a given earth model and its miniature model; namely, if the parameters (layer thicknesses and S-wave velocities) of the given earth model are shrunk at a certain scale, the phase velocities on the fundamental dispersion curve will decrease at the same scale. Taking advantage of this relation, we propose a broad learning framework for search space design in a data-driven manner. First, a training set of minified models is generated using the forward calculation of dispersion curves. Then, a mapping relationship between dispersion curves and minified earth models is built via a broad learning network. Finally, a search space for the actual earth model is designed using the minified model based on the relation. As the ranges of parameters in the minified model are much smaller, the network can find the model parameters for the corresponding dispersion curve quickly and easily. Numerical simulations and field data applications demonstrate the reliability and effectiveness of the proposed method for search space design. It is concluded that the proposed method can promote accurate estimation of S-wave velocities by Rayleigh wave inversion without experience-based judgment.

Source-independent time-lapse full-waveform inversion for anisotropic media

Time-lapse full-waveform inversion can provide high-resolution information about changes in the reservoir properties during hydrocarbon production and CO2 injection. However, the accuracy of the estimated source wavelet, which is critically important for time-lapse FWI, is often insufficient for field-data applications. The so-called “source-independent” FWI is designed to reduce the influence of the source wavelet on the inversion results. We incorporate the convolution-based source-independent technique into a time-lapse FWI algorithm for VTI (transversely isotropic with a vertical symmetry axis) media. The gradient of the modified FWI objective function is obtained from the adjoint-state method. The algorithm is tested on a model with a graben structure and the modified VTI Marmousi model using three time-lapse strategies (the parallel-difference, sequential-difference, and double-difference methods). The results confirm the ability of the developed methodology to reconstruct the localized time-lapse parameter variations even for a strongly distorted source wavelet. The algorithm remains robust in the presence of moderate noise in the input data but the accuracy of the estimated time-lapse changes depends on the model complexity.

Characterization of a Volcanic Gas Reservoir Using Seismic Dispersion and Fluid Mobility Attributes

Abstract Seismic dispersion and fluid mobility attributes are used to characterize a volcanic gas reservoir in the Songliao Basin of China. A rock physics model is constructed to describe poroelastic behaviors associated with heterogeneous fluids saturation within the volcanic gas reservoirs, where velocity dispersion and attenuation of propagating waves are attributed to the wave-induced fluid flow described by the patchy saturation theory. Modeling results indicate that the frequency-dependent bulk modulus at the seismic frequency is more sensitive to gas saturation than the P-wave velocity dispersion. Accordingly, a new inversion method is developed to compute bulk-modulus-related dispersion attribute DK for improved characterization of volcanic gas reservoirs. Synthetic tests indicate that DK is more sensitive than traditional P-wave dispersion attribute DP to the variations of reservoir properties. The high value of dispersion attribute DK indicates the volcanic gas reservoirs with high porosity and gas saturation. At the same time, fluid mobility attribute FM can discriminate the volcanic gas reservoir as DK. Field data applications illustrate that DK and FM exhibit anomalies to the gas zones in the volcanic gas reservoir on the cross-well section. However, DK is more robust than FM to identify favorable zones on horizontal slices for specific target layers. Overall, rock physical modeling provides insights into the poroelastic behaviors of volcanic gas reservoirs, and inversion for seismic dispersion attribute DK improves hydrocarbon detection in the volcanic gas reservoir.