Sparsity-promoting Inversion Research Articles

Deep preconditioners and their application to seismic wavefield processing

Seismic data processing heavily relies on the solution of physics-driven inverse problems. In the presence of unfavourable data acquisition conditions (e.g., regular or irregular coarse sampling of sources and/or receivers), the underlying inverse problem becomes very ill-posed and prior information is required to obtain a satisfactory solution. Sparsity-promoting inversion, coupled with fixed-basis sparsifying transforms, represent the go-to approach for many processing tasks due to its simplicity of implementation and proven successful application in a variety of acquisition scenarios. Nevertheless, such transforms rely on the assumption that seismic data can be represented as a linear combination of a finite number of basis functions. Such an assumption may not always be fulfilled, thus producing sub-optimal solutions. Leveraging the ability of deep neural networks to find compact representations of complex, multi-dimensional vector spaces, we propose to train an AutoEncoder network to learn a nonlinear mapping between the input seismic data and a representative latent manifold. The trained decoder is subsequently used as a nonlinear preconditioner for the solution of the physics-driven inverse problem at hand. Through synthetic and field data examples, the proposed nonlinear, learned transformations are shown to outperform fixed-basis transforms and converge faster to the sought solution for a variety of seismic processing tasks, ranging from deghosting to wavefield separation with both regularly and irregularly subsampled data.

The potential of self-supervised networks for random noise suppression in seismic data

Noise suppression is an essential step in many seismic processing workflows. A portion of this noise, particularly in land datasets, presents itself as random noise. In recent years, neural networks have been successfully used to denoise seismic data in a supervised fashion. However, supervised learning always comes with the often unachievable requirement of having noisy-clean data pairs for training. Using blind-spot networks, we redefine the denoising task as a self-supervised procedure where the network uses the surrounding noisy samples to estimate the noise-free value of a central sample. Based on the assumption that noise is statistically independent between samples, the network struggles to predict the noise component of the sample due to its randomicity, whilst the signal component is accurately predicted due to its spatio-temporal coherency. Illustrated on synthetic examples, the blind-spot network is shown to be an efficient denoiser of seismic data contaminated by random noise with minimal damage to the signal; therefore, providing improvements in both the image domain and down-the-line tasks, such as post-stack inversion. To conclude our study, the suggested approach is applied to field data and the results are compared with two commonly used random denoising techniques: FX-deconvolution and sparsity-promoting inversion by Curvelet transform. By demonstrating that blind-spot networks are an efficient suppressor of random noise, we believe this is just the beginning of utilising self-supervised learning in seismic applications. • Application of self-supervised deep learning for random noise suppression. • Blind-spot networks shown to effectively reduce white and coloured noise. • SNR improvements seen in the image, frequency and acoustic impedance domains. • Network successfully trained and applied to post-stack field data. • Blind-spot approach comparable to commonly used noise suppression procedures.

Handling gaps in acquisition geometries — Improving Marchenko-based imaging using sparsity-promoting inversion and joint inversion of time-lapse data

Marchenko focusing and imaging are novel methods for correctly handling multiple scattered energy while processing seismic data. However, strict requirements in the acquisition geometry, specifically the colocation of sources and receivers as well as dense and regular sampling, currently constrain their practical applicability. We have reformulated the Marchenko equations to handle the case in which there are gaps in the source geometry while the receiver sampling remains regular (or the opposite, by means of reciprocity). Using synthetic data based on a velocity model that produces strong interbed multiples, we test different solvers for the newly formulated inversion problem, and we compare these results to results obtained by applying standard Marchenko inversion to a previously reconstructed data set. When using the unreconstructed data set, the ability of the Marchenko equations to retrace multiple reflected energy deteriorates. Sparsity-promoting Marchenko inversion, while improving the appearance of focusing functions, barely decreases multiple leakage in gathers and does not visibly improve the final image when compared to standard least-squares inversion. On the other hand, reconstructing the wavefield in advance restores the proper functioning of the Marchenko methods. Further, we test a joint inversion technique designed for time-lapse data with nonrepeated geometries and originally intended to be solved using sparsity-promoting inversion. Motivated by our previous results, we compare images produced by this method to images produced by solving the same joint inversion problem without sparsity constraint. We find that the joint inversion alone hardly improves the resulting images but, when combined with the sparsity constraint, it leads to better noise and multiple suppression and produces a clean time-lapse image. Overall, none of the results from sparsity-promoting inversion techniques match the results obtained when reconstructing the wavefield in advance. We show that this can be explained by the comparatively slow convergence rate of the sparsity-promoting Marchenko inversion.

Separation of prestack seismic diffractions using an improved sparse apex-shifted hyperbolic Radon transform

The apex-shifted hyperbolic Radon transform (ASHRT) which is defined as an extension of standard hyperbolic Radon transform (HRT) shifts the apexes of basis functions along the offset. In this paper, we develop an improved sparse ASHRT to separate the diffractions from reflections before stacking. To speed up the calculation, the forward and adjoint operators of ASHRT in the time domain are replaced with the Fourier-kernel Stolt-based modelling and imaging operators. To overcome the limitation of velocity variations, the time axis stretching is implemented. With lower computation cost compared to the time domain operators, we can collapse the apex-shifted events with one velocity. Meanwhile, we introduce a sparsity-promoting inversion by the fast iterative shrinkage thresholding algorithm (FISTA), which produces a sparse Radon panel to separate the diffracted energy easily. Two synthetic examples of variable velocity model show that our proposition is robust and efficient. Another marine data example further demonstrates the effectiveness of this method in separation of prestack seismic diffractions.

A multi-sensor data-driven methodology for all-sky passive microwave inundation retrieval

Abstract. We present a multi-sensor Bayesian passive microwave retrieval algorithm for flood inundation mapping at high spatial and temporal resolutions. The algorithm takes advantage of observations from multiple sensors in optical, short-infrared, and microwave bands, thereby allowing for detection and mapping of the sub-pixel fraction of inundated areas under almost all-sky conditions. The method relies on a nearest-neighbor search and a modern sparsity-promoting inversion method that make use of an a priori dataset in the form of two joint dictionaries. These dictionaries contain almost overlapping observations by the Special Sensor Microwave Imager and Sounder (SSMIS) on board the Defense Meteorological Satellite Program (DMSP) F17 satellite and the Moderate Resolution Imaging Spectroradiometer (MODIS) on board the Aqua and Terra satellites. Evaluation of the retrieval algorithm over the Mekong Delta shows that it is capable of capturing to a good degree the inundation diurnal variability due to localized convective precipitation. At longer timescales, the results demonstrate consistency with the ground-based water level observations, denoting that the method is properly capturing inundation seasonal patterns in response to regional monsoonal rain. The calculated Euclidean distance, rank-correlation, and also copula quantile analysis demonstrate a good agreement between the outputs of the algorithm and the observed water levels at monthly and daily timescales. The current inundation products are at a resolution of 12.5 km and taken twice per day, but a higher resolution (order of 5 km and every 3 h) can be achieved using the same algorithm with the dictionary populated by the Global Precipitation Mission (GPM) Microwave Imager (GMI) products.

Seismic data restoration with a fast L1norm trust region method

Seismic data restoration is a major strategy to provide reliable wavefield when field data dissatisfy the Shannon sampling theorem. Recovery by sparsity-promoting inversion often get sparse solutions of seismic data in a transformed domains, however, most methods for sparsity-promoting inversion are line-searching methods which are efficient but are inclined to obtain local solutions. Using trust region method which can provide globally convergent solutions is a good choice to overcome this shortcoming. A trust region method for sparse inversion has been proposed, however, the efficiency should be improved to suitable for large-scale computation. In this paper, a new L1 norm trust region model is proposed for seismic data restoration and a robust gradient projection method for solving the sub-problem is utilized. Numerical results of synthetic and field data demonstrate that the proposed trust region method can get excellent computation speed and is a viable alternative for large-scale computation.

Prior model identification during subsurface flow data integration with adaptive sparse representation techniques

Construction of predictive reservoir models invariably involves interpretation and interpolation between limited available data and adoption of imperfect modeling assumptions that introduce significant subjectivity and uncertainty into the modeling process. In particular, uncertainty in the geologic continuity model can significantly degrade the quality of fluid displacement patterns and predictive modeling outcomes. Here, we address a standing challenge in flow model calibration under uncertainty in geologic continuity by developing an adaptive sparse representation formulation for prior model identification (PMI) during model calibration. We develop a flow-data-driven sparsity-promoting inversion to discriminate against distinct prior geologic continuity models (e.g., variograms). Realizations of reservoir properties from each geologic continuity model are used to generate sparse geologic dictionaries that compactly represent models from each respective prior. For inversion initially the same number of elements from each prior dictionary is used to construct a diverse geologic dictionary that reflects a wide range of variability and uncertainty in the prior continuity. The inversion is formulated as a sparse reconstruction problem that inverts the flow data to identify and linearly combine the relevant elements from the large and diverse set of geologic dictionary elements to reconstruct the solution. We develop an adaptive sparse reconstruction algorithm in which, at every iteration, the contribution of each dictionary to the solution is monitored to replace irrelevant (insignificant) elements with more geologically relevant (significant) elements to improve the solution quality. Several numerical examples are used to illustrate the effectiveness of the proposed approach for identification of geologic continuity in practical model calibration problems where the uncertainty in the prior geologic continuity model can lead to biased inversion results and prediction.

Nonequispaced curvelet transform for seismic data reconstruction: A sparsity-promoting approach

We extend our earlier work on the nonequispaced fast discrete curvelet transform (NFDCT) and introduce a second generation of the transform. This new generation differs from the previous one by the approach taken to compute accurate curvelet coefficients from irregularly sampled data. The first generation relies on accurate Fourier coefficients obtained by an [Formula: see text]-regularized inversion of the nonequispaced fast Fourier transform (FFT) whereas the second is based on a direct [Formula: see text]-regularized inversion of the operator that links curvelet coefficients to irregular data. Also, by construction the second generation NFDCT is lossless unlike the first generation NFDCT. This property is particularly attractive for processing irregularly sampled seismic data in the curvelet domain and bringing them back to their irregular record-ing locations with high fidelity. Secondly, we combine the second generation NFDCT with the standard fast discrete curvelet transform (FDCT) to form a new curvelet-based method, coined nonequispaced curvelet reconstruction with sparsity-promoting inversion (NCRSI) for the regularization and interpolation of irregularly sampled data. We demonstrate that for a pure regularization problem the reconstruction is very accurate. The signal-to-reconstruction error ratio in our example is above [Formula: see text]. We also conduct combined interpolation and regularization experiments. The reconstructions for synthetic data are accurate, particularly when the recording locations are optimally jittered. The reconstruction in our real data example shows amplitudes along the main wavefronts smoothly varying with limited acquisition imprint.

Simply denoise: Wavefield reconstruction via jittered undersampling

We present a new, discrete undersampling scheme designed to favor wavefield reconstruction by sparsity-promoting inversion with transform elements localized in the Fourier domain. The work is motivated by empirical observations in the seismic community, corroborated by results from compressive sampling, that indicate favorable (wavefield) reconstructions from random rather than regular undersampling. Indeed, random undersampling renders coherent aliases into harmless incoherent random noise, effectively turning the interpolation problem into a much simpler denoising problem. A practical requirement of wavefield reconstruction with localized sparsifying transforms is the control on the maximum gap size. Unfortunately, random undersampling does not provide such a control. Thus, we introduce a sampling scheme, termed jittered undersampling, that shares the benefits of random sampling and controls the maximum gap size. The contribution of jittered sub-Nyquist sampling is key in formu-lating a versatile wavefield sparsity-promoting recovery scheme that follows the principles of compressive sampling. After the behavior of the jittered-undersampling scheme in the Fourier domain is analyzed, its performance is studied for curvelet recovery by sparsity-promoting inversion (CRSI). The findings on synthetic and real seismic data indicate an improvement of several decibels over recovery from regularly undersampled data for the same amount of data collected.

Non-parametric seismic data recovery with curvelet frames

SUMMARY Seismic data recovery from data with missing traces on otherwise regular acquisition grids forms a crucial step in the seismic processing flow. For instance, unsuccessful recovery leads to imaging artefacts and to erroneous predictions for the multiples, adversely affecting the performance of multiple elimination. A non-parametric transform-based recovery method is presented that exploits the compression of seismic data volumes by recently developed curvelet frames. The elements of this transform are multidimensional and directional and locally resemble wave fronts present in the data, which leads to a compressible representation for seismic data. This compression enables us to formulate a new curvelet-based seismic data recovery algorithm through sparsity-promoting inversion. The concept of sparsity-promoting inversion is in itself not new to geophysics. However, the recent insights from the field of ‘compressed sensing’ are new since they clearly identify the three main ingredients that go into a successful formulation of a recovery problem, namely a sparsifying transform, a sampling strategy that subdues coherent aliases and a sparsity-promoting program that recovers the largest entries of the curvelet-domain vector while explaining the measurements. These concepts are illustrated with a stylized experiment that stresses the importance of the degree of compression by the sparsifying transform. With these findings, a curvelet-based recovery algorithms is developed, which recovers seismic wavefields from seismic data volumes with large percentages of traces missing. During this construction, we benefit from the main three ingredients of compressive sampling, namely the curvelet compression of seismic data, the existence of a favourable sampling scheme and the formulation of a large-scale sparsity-promoting solver based on a cooling method. The recovery performs well on synthetic as well as real data by virtue of the sparsifying property of curvelets. Our results are applicable to other areas such as global seismology.

Multiple prediction from incomplete data with the focused curvelet transform

Incomplete data represents a major challenge for a successful prediction and subsequent removal of multiples. In this paper, a new method will be represented that tackles this challenge in a two-step approach. During the first step, the recenly developed curvelet-based recovery by sparsity-promoting inversion (CRSI) is applied to the data, followed by a prediction of the primaries. During the second high-resolution step, the estimated primaries are used to improve the frequency content of the recovered data by combining the focal transform, defined in terms of the estimated primaries, with the curvelet transform. This focused curvelet transform leads to an improved recovery, which can subsequently be used as input for a second stage of multiple prediction and primary-multiple separation.