The Radon transform (RT) suffers from the typical problems of loss of resolution and aliasing that arise as a consequence of incomplete information, including limited aperture and discretization. Sparseness in the Radon domain is a valid and useful criterion for supplying this missing information, equivalent somehow to assuming smooth amplitude variation in the transition between known and unknown (missing) data. Applying this constraint while honoring the data can become a serious challenge for routine seismic processing because of the very limited processing time available, in general, per common midpoint. To develop methods that are robust, easy to use and flexible to adapt to different problems we have to pay attention to a variety of algorithms, operator design, and estimation of the hyperparameters that are responsible for the regularization of the solution.In this paper, we discuss fast implementations for several varieties of RT in the time and frequency domains. An iterative conjugate gradient algorithm with fast Fourier transform multiplication is used in all cases. To preserve the important property of iterative subspace methods of regularizing the solution by the number of iterations, the model weights are incorporated into the operators. This turns out to be of particular importance, and it can be understood in terms of the singular vectors of the weighted transform. The iterative algorithm is stopped according to a general cross validation criterion for subspaces. We apply this idea to several known implementations and compare results in order to better understand differences between, and merits of, these algorithms.

It is well known that a sparse hyperbolic Radon transform (RT) can be used to extend the aperture of aperture limited data, filter noise, and fill gaps. In the same manner, an elliptical RT can achieve similar results when applied to slant stack sections. A problem with these transformations is that they have a time‐variant kernel that results in slow implementation. By defining a model space in terms of an irregularly sampled velocity space to minimize the number of unknowns during the inversion and using sparse matrices, however, the computation time can be kept low enough for practical application. We implement hyperbolic and elliptical time domain RTs by using inversion via weighted conjugate gradient methods with a sparseness constraint. The hyperbolic RT performs accurate interpolation in common‐midpoint (CMP) gathers, while the elliptical RT attenuates sampling artifacts in slant stack sections obtained from CMP gathers with poor sampling and gaps.

Although 3D seismic data are being acquired in larger volumes than ever before, the spatial sampling of these volumes is not always adequate for certain seismic processes. This is especially true of marine and land wide-azimuth acquisitions, leading to the development of multidimensional data interpolation techniques. Simultaneous interpolation in all five seismic data dimensions ͑inline, crossline, offset, azimuth, and frequency͒ has great utility in predicting missing data with correct amplitude and phase variations. Although there are many techniques that can be implemented in five dimensions, this study focused on sparse Fourier reconstruction. The success of Fourier interpolation methods depends largely on two factors: ͑1͒ having efficient Fourier transform operators that permit the use of large multidimensional data windows and ͑2͒ constraining the spatial spectrum along dimensions where seismic amplitudes change slowly so that the sparseness and band limitation assumptions remain valid. Fourier reconstruction can be performed when enforcing a sparseness constraint on the 4D spatial spectrum obtained from frequency slices of five-dimensional windows. Binning spatial positions into a fine 4D grid facilitates the use of the FFT, which helps on the convergence of the inversion algorithm. This improves the results and computational efficiency. The 5D interpolation can successfully interpolate sparse data, improve AVO analysis, and reduce migration artifacts. Target geometries for optimal interpolation and regularization of land data can be classified in terms of whether they preserve the original data and whether they are designed to achieve surface or subsurface consistency.

A hyperbolic Radon transform (RT) can be applied with success to attenuate or interpolate hyperbolic events in seismic data. However, this method fails when the hyperbolic events have apexes located at nonzero offset positions. A different RT operator is required for these cases, an operator that scans for hyperbolas with apexes centered at any offset. This procedure defines an extension of the standard hyperbolic RT with hyperbolic basis functions located at every point of the data gather. The mathematical description of such an operator is basically similar to a kinematic poststack time‐migration equation, with the horizontal coordinate being not midpoint but offset. In this paper, this transformation is implemented by using a least‐squares conjugate gradient algorithm with a sparseness constraint. Two different operators are considered, one in the time domain and the other in the frequency‐wavenumber domain (Stolt operator). The sparseness constraint in the time‐offset domain is essential for resampling and for interpolation. The frequency‐wavenumber domain operator is very efficient, not much more expensive in computation time than a sparse parabolic RT, and much faster than a standard hyperbolic RT. Examples of resampling, interpolation, and coherent noise attenuation using the frequency‐wavenumber domain operator are presented. Near and far offset gaps are interpolated in synthetic and real shot gathers, with simultaneous resampling beyond aliasing. Waveforms are well preserved in general except when there is little coherence in the data outside the gaps or events with very different velocities are located at the same time. Multiples of diffractions are predicted and attenuated by subtraction from the data.

Deep-learning techniques appear to be poised to play very important roles in our processing flows for inversion and interpretation of seismic data. The most successful seismic applications of these complex pattern-identifying networks will, presumably, be those that also leverage the deterministic physical models on which we normally base our seismic interpretations. If this is true, algorithms belonging to theory-guided data science, whose aim is roughly this, will have particular applicability in our field. We have developed a theory-designed recurrent neural network (RNN) that allows single- and multidimensional scalar acoustic seismic forward-modeling problems to be set up in terms of its forward propagation. We find that training such a network and updating its weights using measured seismic data then amounts to a solution of the seismic inverse problem and is equivalent to gradient-based seismic full-waveform inversion (FWI). By refining these RNNs in terms of optimization method and learning rate, comparisons are made between standard deep-learning optimization and nonlinear conjugate gradient and limited-memory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS) optimized algorithms. Our numerical analysis indicates that adaptive moment (or Adam) optimization with a learning rate set to match the magnitudes of standard FWI updates appears to produce the most stable and well-behaved waveform inversion results, which is reconfirmed by a multidimensional 2D Marmousi experiment. Future waveform RNNs, with additional degrees of freedom, may allow optimal wave propagation rules to be solved for at the same time as medium properties, reducing modeling errors.

A method is described for filtering magnetotelluric (MT) data in the wavelet domain that requires a minimum of human intervention and leaves good data sections unchanged. Good data sections are preserved because data in the wavelet domain is analyzed through hierarchies, or scale levels, allowing separation of noise from signals. This is done without any assumption on the data distribution on the MT transfer function. Noisy portions of the data are discarded through thresholding wavelet coefficients. The procedure can recognize and filter out point defects that appear as a fraction of unusual observations of impulsive nature either in time domain or frequency domain. Two examples of real MT data are presented, with noise caused by both meteorological activity and power-line contribution. In the examples given in this paper, noise is better seen in time and frequency domains, respectively. Point defects are filtered out to eliminate their deleterious influence on the MT transfer function estimates. After the filtering stage, data is processed in the frequency domain, using a robust algorithm to yield two sets of reliable MT transfer functions.

We present a technique for deblending seismic data acquired with simultaneous shots. We propose the apex shifted Radon transform (ASRT) method to collapse energy from different shots to their apexes or shot locations. The technique is more suited for streamer acquisition, although other uses are possible. The ASRT is implemented via a Stolt migration operator with a high resolution least squares algorithm. The transform collapses all events to their apexes independently of their geometry (offset and azimuth) and does not require the knowledge of shot locations or velocities. It is as fast as a 2D Radon transform and therefore can be applied during acquisition as a preliminary QC.

No abstract