Unblended Data Research Articles

Free-surface multiple attenuation and seismic deghosting for blended data using convolutional neural networks

Simultaneous source acquisition has become common over the past few decades for marine seismic surveys because of the increased efficiency of seismic acquisition by limiting the time, reducing the cost, and having less environmental impact than conventional single-source (or unblended) acquisition surveys. For simultaneous source acquisition, seismic sources at different locations are fired with time delays, and the recorded data are referred to as the blended data. The air-water interface (or free surface) creates strong multiples and ghost reflections for blended seismic data. The multiples and/or ghost reflections caused by a source in the blended data overlap with the primary reflections of another source, thus creating a strong interference between the primary and multiple events of different sources. We develop a convolutional neural network (CNN) method to attenuate free-surface multiples and remove ghost reflections simultaneously from the blended seismic data. The CNN-based solution that we develop operates on single traces and is not sensitive to the missing near-offset traces, missing traces, and irregular/sparse acquisition parameters (e.g., for ocean-bottom node acquisition and time-lapse monitoring studies). We illustrate the efficacy of our free-surface multiple attenuation and seismic deghosting method by presenting synthetic and field data applications. The numerical experiments demonstrate that our CNN-based approach for simultaneously attenuating free-surface multiples and removing ghost reflections can be applied to the blended data without the deblending step. Although the interference of primaries and multiples from different shots in the blended data makes free-surface multiple attenuation harder than the unblended data, we determine that our CNN-based method effectively attenuates free-surface multiples in the blended synthetic and field data even when the delay time for the blending is different in the training data than in the data to which the CNN is applied.

Iterative deblending using unsupervised learning with double-deep neural networks

Simultaneous source acquisition technology can greatly improve seismic acquisition efficiency. However, due to continuous shooting and serious crosstalk noise of the adjacent sources in seismic data, simultaneous source data cannot be directly used in conventional data processing procedures. Therefore, simultaneous source data need to be deblended to obtain the conventional shot record. Under densely sampled sources, we have developed a novel unsupervised deep learning (UDL) method based on the double-deep neural networks for iterative inversion deblending of simultaneous source data. Our UDL, which is mainly composed of the residual neural network (R-net) and the U-net neural network, has excellent nonlinear optimization ability. The total loss function design can optimize our UDL in the correct direction and avoid the problem of overfitting. By minimizing the total loss function, the R-net and U-net branches of the UDL can extract the coherent effective signals of all sources and suppress the crosstalk noise. The most prominent advantage of our UDL method is that it does not require label data, and the training data set does not contain raw unblended data, thus solving the problem of missing training data sets. The examples with two synthetic and one field data set are used to prove the effectiveness of iterative inversion deblending of simultaneous source data based on our UDL method when sources are within a small distance of each other. By comparing our UDL method with the traditional curvelet-based and contourlet-based methods, the superiority of our method in the quality of separation results is demonstrated.

A multi‐data training method for a deep neural network to improve the separation effect of simultaneous‐source data

AbstractWithin the field of seismic data acquisition with active sources, the technique of acquiring simultaneous data, also known as blended data, offers operational advantages. The preferred processing of blended data starts with a step of deblending, that is separation of the data acquired by the different sources, to produce data that mimic data from a conventional seismic acquisition and can be effectively processed by standard methods. Recently, deep learning methods based on the deep neural network have been applied to the deblending task with promising results, in particular using an iterative approach. We propose an enhancement to deblending with an iterative deep neural network, whereby we modify the training stage of the deep neural network in order to achieve better performance through the iterations. We refer to the method that only uses the blended data as the input data as the general training method. Our new multi‐data training method allows the deep neural network to be trained by the data set with the input patches composed of blended data, noisy data with low amplitude crosstalk noise, and unblended data, which can improve the ability of the deep neural network to remove crosstalk noise and protect weak signal. Based on such an extended training data set, the multi‐data training method embedded in the iterative separation framework can result in different outputs at different iterations and converge to the best result in a shorter iteration number. Transfer learning can further improve the generalization and separation efficacy of our proposed method to deblend the simultaneous‐source data. Our proposed method is tested on two synthetic data and two field data to prove the effectiveness and superiority in the deblending of the simultaneous‐source data compared with the general training method, generic noise attenuation network and low‐rank matrix factorization methods.

Joint deblending and source time inversion

The simultaneous source technique has been widely applied in seismic field acquisition and has achieved great success in the past few decades. Many studies have shown that the inversion-based source-separation algorithms are more robust than filtering-based methods. However, inversion-based methods depend on accurate shot times. Here, we tackle the long-standing issue of the shot times being inaccurate or unavailable during iterative inversion by proposing a joint inversion framework to simultaneously separate the blended sources and invert for the shot time. We formulate a nonlinear inverse problem that contains two unknowns, i.e., the unblended data and the shot-time vector, and we develop a Gauss-Newton method to iteratively invert the shot-time vector given an estimate of the unblended data. Then, the estimated shot-time vector is fixed for iterative source separation following a traditional deblending framework. The two aforementioned steps are recursively implemented until they converge or reach a maximum number of iterations. We examine the proposed method through several synthetic and field data examples. Results indicate that the proposed joint inversion framework is effective and the low-frequency component matters during the nonlinear inversion of the shot-time vector.

An unsupervised learning approach to deblend seismic data from denser shot coverage surveys

SUMMARY The simultaneous source data obtained by simultaneous source acquisition contain crosstalk noise and cannot be directly used in conventional data processing procedures. Therefore, it is necessary to deblend the blended wavefield to obtain the conventionally acquired single-shot recordings. In this study, we propose an iterative inversion method based on the unsupervised deep neural network (UDNN) to deblend the simultaneous source data from a denser shot coverage survey (DSCS). In the common receiver gather (CRG), the coherent effective signals in the blended data of the primary and secondary sources are similar. We exploit the excellent nonlinear optimization capability of the U-net network to extract similar coherent signals from the blended data of the primary and secondary sources by minimizing the total loss function. The proposed UDNN method does not need to use the raw unblended data as label data, which solves the problem of missing label data and is suitable for deblending the simultaneous source data in different work areas with complex underground structures. One synthetic data and one field data examples are used to prove that the proposed method can suppress crosstalk noise and protect weak effective signals effectively, and achieve good effectiveness for the separation of simultaneous source data.

Deblending based on the frequency-structure weighted threshold: The case of the slope-constrained f-k transform

In the process of separating blended data, conventional methods based on sparse inversion assume that the primary source is coherent and the secondary source is randomized. The L1-norm, the commonly used regularization term, uses a global threshold (GT) to process the sparse spectrum in the transform domain; however, when the threshold is relatively high, more high-frequency information from the primary source will be lost. For this reason, we have analyzed the generation principle of blended data based on the convolution theory and then conclude that the blended data are only randomly distributed in the spatial domain. Taking the slope-constrained frequency-wavenumber ( f- k) transform as an example, we adopt a frequency-dependent threshold, which reduces the high-frequency loss during the deblending process. Then, we use a structure-weighted threshold (SWT) in which the energy from the primary source is concentrated along the wavenumber direction. The combination of frequency and SWTs effectively improves the deblending performance. Model and field data find that our frequency-structure weighted threshold has better frequency preservation than the GT. The weighted threshold can better retain the high-frequency information of the primary source, and the similarity between other frequency-band data and the unblended data has been improved.

Acquisition/Processing: Machine learning-based deblending: Dispersed source array data example

Machine learning (ML) has proven its value in the seismic industry with successful implementations in areas of seismic interpretation such as fault and salt dome detection and velocity picking. The field of seismic processing research also is shifting toward ML applications in areas such as tomography, demultiple, and interpolation. Here, a supervised ML deblending algorithm is illustrated on a dispersed source array (DSA) data example in which both high- and low-frequency vibrators were deployed simultaneously. Training data pairs of blended and corresponding unblended data were constructed from conventional (unblended) data from another survey. From this training data, the method can automatically learn a deblending operator that is used to deblend for both the low- and the high-frequency vibrators of the DSA data. The results obtained on the DSA data are encouraging and show that the ML deblending method can offer a good performing, less user-intensive alternative to existing deblending methods.

Advanced Deblending Scheme for Independent Simultaneous Source Data

Independent Simultaneous Source (ISS®) technology is an attractive way to provide high acquisition production rates and affordable higher density seismic imaging. In this paper, we propose an advanced deblending scheme to address the source separation requirement of ISS® data. The deblending scheme includes a) 3D low-rank decomposition based iterative noise modelling, and subtraction, b) signal-to-noise ratio map guided residual denoise and c) erratic noise attenuation. The application is tested on actual shallow water OBC dataset, which was pseudo-blended using a realistic dual vessel acquisition plan. The result shows that this advanced deblending scheme can successfully recover individual source responses. The deblended data have a minimal difference compared with the actual unblended data.

Velocity analysis of simultaneous-source data using high-resolution semblance—coping with the strong noise

Direct imaging of simultaneous-source (or blended) data, without the need of deblending, requires a precise subsurface velocity model. In this paper, we focus on the velocity analysis of simultaneous-source data using the normal moveout-based velocity picking approach.We demonstrate that it is possible to obtain a precise velocity model directly from the blended data in the common-midpoint domain. The similarity-weighted semblance can help us obtain much better velocity spectrum with higher resolution and higher reliability compared with the traditional semblance. The similarity-weighted semblance enforces an inherent noise attenuation solely in the semblance calculation stage, thus it is not sensitive to the intense interference. We use both simulated synthetic and field data examples to demonstrate the performance of the similarity-weighted semblance in obtaining reliable subsurface velocity model for direct migration of simultaneous-source data. The migrated image of blended field data using prestack Kirchhoff time migration approach based on the picked velocity from the similarity-weighted semblance is very close to the migrated image of unblended data.

Separation and reconstruction of simultaneous source data via iterative rank reduction

We have developed a rank-reduction algorithm based on singular spectrum analysis (SSA) that is capable of suppressing the interferences generated by simultaneous source acquisition. We evaluated an inversion scheme that minimizes the misfit between predicted and observed blended data in [Formula: see text] domain subject to a low-rank constraint that is applied to data in the [Formula: see text] domain. In particular, we developed an iterative algorithm by adopting the projected gradient method with the SSA filter acting as the projection operator. This method entails extracting small patches of data from a common receiver gather and organizing the spatial data at a given monochromatic frequency into a Hankel matrix. For the ideal unblended data, Hankel matrices extracted from the data are of low rank. The incoherent interferences in common-receiver domain caused by simultaneously fired shots increase the rank of the aforementioned Hankel matrices. Therefore, rank-reduction filtering is an effective way to annihilate source interferences while preserving the unblended signal. Through tests with synthetic examples, we found that the interference can be effectively suppressed by the proposed method. In addition, we found that the proposed algorithm can be modified to simultaneously cope with deblending and data recovery. A real survey acquired in the Gulf of Mexico was used to mimic a simultaneous-source acquisition with missing shot locations. The algorithm was able to recover the missing shot gathers from the blended acquisition with an improvement of the signal quality of approximately 12 dB.

Deblending using normal moveout and median filtering in common-midpoint gathers

The benefits of simultaneous source acquisition are compromised by the challenges of dealing with intense blending noise. In this paper, we propose a processing workflow for blended data. The incoherent property of blending noise in the common-midpoint gathers is utilized for applying median filtering along the spatial direction after normal moveout (NMO) correction. The key step in the proposed workflow is that we need to obtain a precise velocity estimation which is required by the subsequent NMO correction. Because of the intense blending noise, the velocity scan cannot be obtained in one step. We can recursively polish both the deblended result and the velocity estimation by deblending using the updated velocity estimation and velocity scanning using the updated deblended result. We use synthetic and field data examples to demonstrate the performance of the proposed approach. The migrated image of deblended data is cleaner than that of blended data and is similar to that of unblended data.

Separation of seismic blended data by sparse inversion over dictionary learning

Abstract Recent development of blended acquisition calls for the new procedure to process blended seismic measurements. Presently, deblending and reconstructing unblended data followed by conventional processing is the most practical processing workflow. We study seismic deblending by advanced sparse inversion with a learned dictionary in this paper. To make our method more effective, hybrid acquisition and time-dithering sequential shooting are introduced so that clean single-shot records can be used to train the dictionary to favor the sparser representation of data to be recovered. Deblending and dictionary learning with l1-norm based sparsity are combined to construct the corresponding problem with respect to unknown recovery, dictionary, and coefficient sets. A two-step optimization approach is introduced. In the step of dictionary learning, the clean single-shot data are selected as trained data to learn the dictionary. For deblending, we fix the dictionary and employ an alternating scheme to update the recovery and coefficients separately. Synthetic and real field data were used to verify the performance of our method. The outcome can be a significant reference in designing high-efficient and low-cost blended acquisition.

ChemInform Abstract: The Fundamental Torsion Band in Acetaldehyde

Abstract High-resolution spectroscopy and low-temperature experimental conditions allowed us to carry out the rotational analysis of the ν 15 fundamental band of acetaldehyde, observed around 150 cm −1 . Some 1000 lines of A and E types have been assigned to the main band and some 90 A -type lines have been identified in the first overtone of the torsion mode. A simultaneous fit, using the nonrigid internal axis method, of a very severely selected set of unblended FIR data from the fundamental band (214 lines with J ≤ 10 and K V 3 and V 6 of the barrier for internal rotation as well as the rotationless origins of the ν 15 fundamental band and its first overtone were determined with improved accuracy, compared to the data previously available in the literature. Some difficulties in treating v t = 1 microwave data led us to suspect problems in the theoretical model adopted or in the present v t = 1 microwave data set. Finally, a fit of selected unblended data from the first overtone of the torsion (29 A -type lines), together with the previously mentioned selected unblended data from the fundamental band, was carried out but resulted in a standard deviation somewhat higher than was obtained with the fundamental band alone.

The fundamental torsion band in acetaldehyde

High-resolution spectroscopy and low-temperature experimental conditions allowed us to carry out the rotational analysis of the ν 15 fundamental band of acetaldehyde, observed around 150 cm −1. Some 1000 lines of A and E types have been assigned to the main band and some 90 A-type lines have been identified in the first overtone of the torsion mode. A simultaneous fit, using the nonrigid internal axis method, of a very severely selected set of unblended FIR data from the fundamental band (214 lines with J ≤ 10 and K <- 9), together with microwave data published in the literature, allowed us to obtain accurate molecular parameters describing describing internal and overall rotation in acetaldehyde. In particular, the Fourier coefficients V 3 and V 6 of the barrier for internal rotation as well as the rotationless origins of the ν 15 fundamental band and its first overtone were determined with improved accuracy, compared to the data previously available in the literature. Some difficulties in treating v t = 1 microwave data led us to suspect problems in the theoretical model adopted or in the present v t = 1 microwave data set. Finally, a fit of selected unblended data from the first overtone of the torsion (29 A-type lines), together with the previously mentioned selected unblended data from the fundamental band, was carried out but resulted in a standard deviation somewhat higher than was obtained with the fundamental band alone.