Anisotropy and attenuation
In the previous report, SEP 163, I demonstrated the potential of using the second-order pseudo-acoustic anisotropic wave equations for full waveform inversion with simple examples. In this report, I investigate some practical issues that anisotropic full waveform inversion implemented with this system of equations faces. These issues are parameter sensitivity, parameterization, and null space. Firstly, to take into account differences in sensitivity of different anisotropic parameters, I suggest a simple normalization technique. This normalization results in dimensionless parameters that better reflect sensivity to recorded data. Secondly, a number of parameterizations, including stiffness coefficients, velocities, and Thomsen's parameters, are tested on the BP 2007 synthetic model. I found that parameterizations with one velocity and two Thomsen's parameters gave the best results in terms of reduction in the objective function. Lastly, to mitigate the null space problem, I regularize the inversion with steering filters based on the migrated image's dip information.
I extend my previous work about the energy imaging condition for anisotropic reverse-time migration to tilted-transverse isotropic media. Synthetic examples show the feasibility of this implementation.
Application of wave-equation migration velocity and Q analysis to the field data from the North Sea (pdf)
The Dolphin Geophysical’s (Dolphin) multi-client field data acquired in the North Sea used in this study has attenuation problems. The area was under the influence of salt tectonics, producing two diapirs. Dolphin interpreted a gas chimney above one diapir, and a channel above the other. The gas chimney forms a migration pathway for the gas to leak and then to accumulate at the shallow position. The shallow gas gives rise to strong attenuation and low interval velocities in the gas area. The channel also has low velocities, and strong attenuation is associated with it. The objective of my study is to update the provided velocity model, especially at the gas and channel area, and to invert for the Q models to recover these two anomalies. Angle domain common image gathers after migration with the current interval velocity show that most of the events are curved down, indicating the current velocity is too high. Furthermore, both the migrated image and the angle gathers show that the events between 26,000 meters (m) to 28,000 m and 38,000 m to 42,000 m are wiped out below the two salt bodies. The attenuation anomalies above two salt bodies are the main reasons for the wiped-out image below. I first applied wave-equation migration velocity analysis to update the current velocity model. As a result, the velocity decreases in the gas and channel area. The angle gathers migrated using the updated velocity model are much more flattened, and the events above the top of salt in the migrated images are more coherent. Then, I applied wave-equation migration Q analysis to invert for the Q models. The estimated Q model shows that the two Q anomalies are recovered and match the interpretation. By using this Q model in seismic migration, I made the seismic events below the anomalies clearly visible, with improved frequency content and coherency of the events.
We use the spectral ratio method to estimate the effect of anelastic attenuation. The proper bandwidth over which to use the method is not constant throughout the entire image of the subsurface. Hand picking the correct bandwidth at every location would be prohibitive for a large 2D, let alone 3D image. Herein, various techniques are explored to systematically identify the upper and lower frequency bounds in order to improve performance of the spectral ratio method over large data sets. These techniques are evaluated with a visual metric to asses the accuracy of their bandwidth picks.
Alejandro Cabrales-Vargas, Biondo Biondi, and Robert Clapp
Linearized Waveform Inversion or least-squares migration is a process that aims at obtaining a better estimation of the subsurface reflectivity, in comparison with conventional migration. During the process, the background model (velocity or slowness) remains invariant. Only the reflectivity component is updated. In this report we revisit the Linearized Waveform Inversion with Velocity Updating theory introduced in a previous report, and present the first synthetic examples. The method introduces controlled updates to the background model during the reflectivity inversion, correcting for slowness inaccuracies that negatively affect seismic amplitudes during conventional linearized waveform inversion. The method incorporates Wave-Equation Migration Velocity Analysis to transform such background model updates into perturbations in the image.
Solving nonlinear inverse problems by linearized model extension - a survey of possible methods (pdf)
Biondo Biondi, Rahul Sarkar, and Joseph Jennings
Nonlinear inverse problems are challenging because gradient-based inverse algorithms may converge toward local minima instead of the desired global minimum. Different methods can be used to solve nonlinear inverse problems based on a linearized model extension; these methods differ in their global-convergence characteristics and in their convergence rate. We present and analyze, both analytically and numerically, three types of such methods. All three methods show attractive global convergence properties. However, our analysis is both incomplete and based on a simple 1D wave-propagation problem where the medium is characterized by a single slowness value. We discuss the convergence rate of the three types of solution we proposed, but, at the current stage of our research, we cannot reach any definitive conclusions on their convergence rate.
Joseph Jennings and Shuki Ronen
The non-uniqueness of simultaneous source deblending inversion is a challenge in simultaneous source separation. We propose to add another constraint to this inversion using radiality, an attribute that can be computed from multicomponent data that are commonly recorded during ocean-bottom node (OBN) or multi-sensor streamer surveys. We describe a simple, proof-of-concept scheme that demonstrates the use of radiality as an additional constraint and show results on an OBN field dataset.
Machine learning and event detection
Fantine Huot and Robert Clapp
Strong localized heterogeneities in the subsurface, such as karst caverns and sinkholes, cause scattering of seismic waves, thereby degrading the images obtained in conventional processing. We explore the possibility of using pattern recognition techniques for detecting these strong heterogeneities from seismic data. Through synthetic models, we generate data with significant scattering. We then perform reverse time migration (RTM) and use various pre-processing techniques to engineer features fit for supervised learning algorithms. Eventually, we use support vector machines (SVM) to classify these features and retrieve the approximate cavern locations.
Ohad Barak and Fantine Huot
In addition to reflection data, seismic recordings contain many different wave modes that are either unwanted or unneeded, degrading overall data quality. We use support vector machines (SVM), a type of supervised learning algorithm, for automatic wave mode classification. We decompose multicomponent translational and rotational seismic data from a field survey into polarization vectors, by applying continuous wavelet transforms (CWT) followed by singular value decomposition (SVD). We train an SVM classifier to distinguish surface waves from body waves from these polarization vectors, and show classification results on different portions of the field data. Our method does not rely on spatial continuity, and can therefore be applied to spatially aliased data.
Ethan Williams and Eileen Martin
We analyze the impact of identifying and removing coherent anthropogenic noise on synthetic Green's functions extracted from ambient noise recorded on a shallow trenched, dense, linear distributed acoustic sensing (DAS) array. Low-cost, low-impact urban seismic surveys are possible with ambient noise recorded by DAS, which uses dynamic strain sensing to detect seismic waves incident to a buried fiber optic cable. However, ambient noise data recorded in urban areas include coherent, time-correlated noise from near-field infrastructure such as cars and trains passing the array, in some cases causing artifacts in estimated Green's functions and yielding potentially incorrect surface wave velocities. Based on our comparison of several methods, we propose an automated, real-time data processing workflow to detect and reduce the impact of these events on data from a dense array in an urban environment. We show the effect of removing such unwanted noise on estimated Green's functions from ambient noise data recorded in Richmond, CA in December 2014 and Fairbanks, AK in August 2015.
Seismic surveys provide us with an abundance of data characteristics. Is there an informative way to visualize a survey's metadata? Using exploratory data analysis (EDA) tools, we investigate a land survey's metadata to detect trends among the observations. We derive indicators of the quality of a seismic trace.
Joseph Jennings and John Washbourne
SEP received a Helmholtz solver during a visit from Chevron researchers. We describe a high-level overview of the workings of the solver and provide simple examples of how this solver may be used for performing full-waveform inversion.
Robert G. Clapp
In order to achieve truly reproducible results the underlying software architecture must be captured. Docker containers, a lightweight alternative to a virtual machine, can be used to snapshot all software dependencies and allow anyone to reproduce an author's results with minimal effort. We demonstrate the effectiveness of Docker containers in several contexts including reproducible research, computer labs, and writing LaTeX documents.