Achieving Robust Compressive Sensing Seismic Acquisition with a Two-Step Sampling Approach † Authors: Anna Titova, Michael B Wakin, Ali C Tura, Nenad Gucunski Published: 2023 ABSTRACT AbstractThe compressive sensing (CS) framework offers a cost-effective alternative to dense alias-free sampling. Designing seismic layouts based on the CS technique imposes the use of specific sampling patterns in addition to the logistical and geophysical requirements. We propose a two-step design process for generating CS-based schemes suitable for seismic applications. During the first step, uniform random sampling is used to generate a random scheme, which is supported theoretically by the restricted isometry property. Following that, designated samples are added to the random scheme to control the maximum distance between adjacent sources (or receivers). The null space property theoretically justifies the additional samples of the second step. Our sampling method generates sampling patterns with a CS theoretical background, controlled distance between adjacent samples, and a flexible number of active and omitted samples. The robustness of two-step sampling schemes for reallocated samples is investigated and CS reconstruction tests are performed. In addition, using this approach, a CS-based 3D seismic survey is designed, and the distributions of traces in fold maps and rose diagrams are analyzed. It is shown that the two-step scheme is suitable for CS-based seismic surveys and field applications.Keywords: signal sampling, survey design, compressive sensing, seismic acquisition 1. Introduction Seismic surveys are designed to balance cost, logistics, and a spatial sampling interval, which is crucial for recording a seismic wavefield. The proper (alias-free) spatial sampling of the seismic wavefield is achieved when the distance between adjacent sources (or receivers) Δd does not exceed the Nyquist sampling criterion ΔdNyq: Δd≤ΔdNyq=Vmin2fmax, where Vmin is the minimum velocity of the media and fmax is the maximum frequency of the source [1]. Depending on Vmin and fmax, the Nyquist criterion may be very small, e.g., 1 m. The sampling theorem can be expanded to relax the Nyquist criterion, but these expansions require measurements of the wavefield gradient [2,3]. Even so, the successful application of the sampling theorem and its expansions may be challenging when the knowledge of Vmin over the area of interest is limited. The alias-free sampling of a seismic wavefield in all spatial directions is rarely satisfied in practice due to the cost constraints of the survey. Various solutions have been adopted and developed to overcome the issue of spatial aliasing. For instance, in typical exploration survey layouts, at least one or two out of four spatial coordinates are densely sampled to facilitate the proper removal of unwanted linear noise [4]. This type of noise is usually represented by the surface waves that are suffering the most from the sparse aliasing sampling intervals. On the data processing side, different algorithms can be applied to mitigate the aliasing in the data, for instance, reconstruction methods based on the prediction error filter [5,6]. In addition to a long-standing dilemma of shooting a dense survey or working with aliased seismic gathers, the ideal regular seismic layout can be compromised by field deployment circumstances: obstacles and areas without access, equipment malfunction, zones with poor recording conditions, logistical constraints, and surface topography. These long-standing, common, and persistent seismic acquisition challenges motivate us to consider non-regular and non-dense survey layouts. Recent advances in signal processing have paved the way for a new cost-effective acquisition paradigm called compressive sensing (CS) [7]. The foundation of CS is the concept of signal sparsity, which is when almost all the energy of a signal is contained in a few transform coefficients of the signal. The key elements of the CS framework include the sampling scheme, sparse transform, and the reconstruction algorithm. Together, they form a framework that can solve undetermined problems, especially those that are common in seismic data processing. One of the earliest works on the application of compressive sensing for seismic signals was published by [8]. The field implementation of CS-based acquisition has been pioneered by [9] and further practiced by [10]. It is worth pointing out that, even before the first applications of the CS framework for the problems of seismic data interpolation, a vast amount of techniques for data reconstruction and dealiasing of spatial wavenumbers has been developed [6]. These techniques are being continuously improved [11]. The CS framework stands out from these methods because it constrains in a direct manner the type of sampling pattern needed for the CS reconstruction workflow to succeed. We focus on CS-based sampling schemes and the application of these schemes for seismic CS. Different CS-based sampling methods can be found in the literature. However, the authors of these methods do not assess the robustness of the prepared CS-based schemes to sample perturbations or reallocations. The latter is an essential consideration for seismic CS-based design since the prepared layout of sources and receivers can be disturbed during the field deployment. In addition, only a few of the proposed CS-based sampling methods are supplemented with complete instructions for generating a proposed sampling scheme. We organize our study as follows. Firstly, we formulate the criteria for CS-based sampling schemes and analyze the strengths and weaknesses of the available sampling methods. Secondly, guided by the theory of the CS framework, we introduce and implement our two-step sampling approach. We evaluate the performance of two-step sampling schemes through CS reconstruction and compare them with the results obtained for the jittered schemes. We also compare the sensitivity of two-step and jittered sampling schemes for sample reallocations. This comparison demonstrates that the two-step scheme is more robust in the face of perturbations in the sampling scheme. A mathematical formulation of the CS framework for seismic applications is provided in Appendix A. Furthermore, in Appendix B, we summarize different tools that can be used to analyze the CS-based schemes. 2. Overview of Sampling Methods Important characteristics of CS-based sampling schemes include the following: CS theoretical guarantees, tolerance to perturbations of scheme samples, and flexibility to incorporate geophysical parameters and logistical constraints. Many methods have been proposed for generating CS-based sampling schemes for seismic applications. We briefly describe these methods by clustering them into five main groups based on sampling mechanisms. The first group of methods includes schemes with random source (receiver) spacing. The random samples may be generated by taking M samples out of N uniformly at random (Mβ, such that each element of the set {1,2,…,α} can be written as the difference of two elements of this set in exactly γ ways. The difference sets may be excellent candidates for CS-based schemes from a mathematical point of view. For example, these sampling schemes achieve the smallest coherence when the Fourier transform is taken as the sparse transform. In addition, all non-DC amplitudes of the Fourier spectrum of a difference set are equal. The disadvantages of using the difference set for seismic survey design start with the absence of control of the distances between adjacent sources (or receivers). Second, the difference set limits the survey parameters, such as the number of sources (or receivers) used in the field, because the difference sets exist only for specific combinations of parameters α, β, and γ. Finally, the difference sets are sensitive to any changes in the scheme: the smallest coherence may no longer be achieved with a few reallocated samples. Such sampling schemes impose additional restrictions on the field implementations of CS-based surveys. The methods in the fourth group are based on a sub-interval strategy [8,33,34,35]. Implementations of this strategy include (i) splitting a source (or receiver) line into sub-intervals and (ii) taking samples within each sub-interval based on a probability distribution. This strategy is devised to control the distance between adjacent sources (or receivers). A recent study by [36] can be used to theoretically investigate the CS reconstruction performance of these methods. One of the disadvantages of sub-interval methods is that repetitive sampling patterns may appear due to the independent choice of the samples in each sub-interval. In addition, the sub-interval strategy may limit the choice of survey parameters. For example, in the jittered sampling, the number of acquired data points M should be a sub-multiple of the number of data points after reconstruction N [8]. In other methods from the sub-interval group [33,35], the relationship between M and N has been made more flexible by randomly varying the number of samples in each subinterval nsub and the length of each subinterval lsub. Consequently, two additional parameters are introduced in the sampling scheme generation process. The relationship between these four parameters (M, N, nsub, and lsub) has to be analyzed to understand how to generate a sampling scheme once specific M and N are given. Low-discrepancy sequences define the fifth group of sampling approaches for building CS-based schemes. In seismic exploration, these sequences were first applied as an optimum discretization scheme for migration algorithms. In particular, ref. [37] used the Halton and Hammersley points to provide a near-optimal discretization for the prestack Kirchhoff migration. As for the CS application, the low-discrepancy sequences may be good candidates because these sequences avoid clusters of sources (or receivers) and large distances between adjacent sources (or receivers). On the other hand, the disadvantages of these sequences include (i) the absence of known CS recovery guarantees and (ii) a logistically challenging implementation of these schemes in the field due to off-grid locations of sources (or receivers). Furthermore, the numerical studies on CS reconstruction with the Hammersley points demonstrate discouraging results [38] and, finally, low-discrepancy sequences may require additional transformation to reduce aliasing [39]. The two-step sampling method was initially proposed by [40]. This paper describes this method in detail and performs different numerical experiments with the two-step sampling schemes. The two-step sampling approach provides sampling schemes with the CS theoretical background, the controlled distance between adjacent sources (or receivers), and the flexible choice of seismic survey parameters. The method combines two types of sampling: random and designated. Random samples are needed to achieve a CS theoretically compliant sampling. The designated samples are purposefully added to the random scheme to control the maximum distance between adjacent sources (or receivers). The random and designated samples are generated sequentially, one after the other. Two-step sampling can be applied in the presence of obstructions in the field and does not limit the choice of seismic survey parameters. In addition, our approach generates schemes that are robust to modifications in a prepared layout of sources (or receivers). The theoretical considerations for the two-step sampling are built based on the Fourier basis. However, the use of this sampling method is not limited to the Fourier sparse transform. In fact, the numerical CS reconstruction experiments are performed with the curvelet sparse transform. 3. Two-Step Sampling Mechanism 3.1. Theoretical Background Two-step sampling merges random and designated samples. The random samples are needed to obtain CS recovery guarantees for the sensing matrix A when the Fourier transform is used as the sparse transform. In this case, the sensing matrix is a member of a special class of matrices—the random partial Fourier matrices. These matrices are built by selecting a subset of M rows uniformly at random from the N rows of the Fourier matrix (M0. The random sampling scheme does not control the distance between adjacent samples, and consequently, the CS reconstruction of signals that are not perfectly sparse may deteriorate. Also, the distance between adjacent samples should be controlled to meet the Fresnel zone requirement of the seismic survey. The designated samples may be added to address this issue. Mathematically, adding designated samples into the random sampling scheme is equivalent to appending new rows into the sensing matrix, which is accomplished by adding new binary rows to the matrix R while keeping the matrix C fixed (Appendix A). These additional rows may change the restricted isometry constant (RIC), and the RIP will no longer hold if the RIC becomes too large, ceasing the CS recovery guarantees. Adding designated samples into the random sampling scheme and, at the same time, preserving the CS recovery guarantees may be achieved by the ℓ2 robust null space property (NSP) (Equation (A8)). First, it is known that sensing matrices with δ2s<4/41 satisfy the ℓ2-robust NSP of order s with 0<ρ<1 and τ>0 depending only on δ2s [12]. Secondly, we demonstrate in Appendix C that the ℓ2 robust NSP stays valid with additional rows appended into the sensing matrix. The combination of the RIP and the NSP provides the schemes with theoretical guarantees and, at the same time, a controlled distance between adjacent samples. 3.2. Implementation The implementation of two-step sampling consists of generating a random scheme and inserting designated samples. We introduce the following notation to facilitate an explanation of our method. Let the set ΩR contain MR random samples obtained during the first step, and let the set ΩD include MD designated samples associated with the second step. The union of these two sets is the set ΩA=ΩR∪ΩD, which contains all samples active in the field. The cardinality of the set ΩA is equal to MA=MR+MD, or in other words, card(MA)=card(MR)+card(MD). The set ΩO denotes a set of omitted samples for which field data are not recorded, and the set Ω=ΩO∪ΩA contains active and omitted samples with card(Ω)=MO+MA=N. We define a gap in a sampling scheme as the distance between adjacent active samples. We consider that samples are placed on a grid, and consequently, the gap can be measured in grid steps. For example, the gap size of g means there are g empty grid steps between adjacent active samples. Firstly, we study how often a random number generator creates random sampling schemes with a small maximum gap. In this experiment, we first generated random sampling schemes with 30% and 50% of active samples, and then we varied the total number of samples in the schemes (N). The histograms of the maximum gap sizes in random schemes with 30% and 50% of active samples are shown in Figure 1a–c for the cases of N=10, N=100, and N=1000, respectively. We used 1,000,000 trials for each case. Figure 1 demonstrates that, as the total number of samples increases, the smallest maximum gap size that appeared over 1,000,000 trials increases. This observation is also supported by Figure 2, which shows the relationship between the smallest maximum gap size in random sampling schemes and the total number of samples. According to Figure 2, when the total number of samples in the scheme exceeds 100, the random sampling schemes with 50% of active samples will have a maximum gap size not smaller than four grid steps. It is worth highlighting that Figure 2 shows the smallest maximum gap size, and the median maximum gap size is larger, as can be observed in Figure 1. Therefore, when the total number of samples is relatively large (e.g., N=500), the small gaps (e.g., gap size of 3) will rarely appear in the schemes with samples chosen uniformly at random. The histograms of the maximum gap sizes in 1,000,000 random schemes with 30% and 50% of active samples for (a) N=10, (b) N=100, and (c) N=1000. With the increase in the total number of samples (active and omitted), the smallest maximum gap size increases. The relationship between the smallest maximum gap size and the total number of samples in random sampling schemes. The maximum gap size exceeds four grid steps once the total number of samples exceeds 100. The designated samples can be added to control the gap size, but at the same time, the total number of samples in the sampling scheme becomes uncontrolled. Analyzing random schemes, one can notice that the gaps in such schemes have different sizes and are distributed differently within the scheme. To achieve the sampling schemes with the fixed number of active samples (MA=MR+MD) and predetermined maximum gap size gmax at the same time, we formulate the following optimization problem: where f(ΩA) is a function that computes the largest gap in the sampling scheme. According to Equation (2), we are interested in finding a realization of random sampling ΩR with the maximum possible number of random samples MR such that the number of all active samples will not exceed MA=MR+MD. The more random samples taken in the first step, the larger the sparsity level s of the RIP. Thus, signals with more non-zero transform coefficients may be reconstructed. Equation (2) outlines three main parameters involved in the search of a random sampling scheme: the maximum gap size gmax, the number of random samples MR, and the number of active samples MA. We implemented the following steps to numerically study the space of these three parameters: (i)Generate a random sampling scheme with M˜R samples; Generate a random sampling scheme with M˜R samples; (ii)Add as many samples as needed to fill in all gaps larger than the gap threshold g˜max; Add as many samples as needed to fill in all gaps larger than the gap threshold g˜max; (iii)Count the number of samples M˜A in the resulting scheme. Count the number of samples M˜A in the resulting scheme. Steps (i)–(iii) are repeated 100 times for each pair M˜R and g˜max, where M˜R/N∈[0.1:0.02:0.9] and g˜max∈[2:1:8]. As for M˜A, this parameter is not constrained and may be different for a given pair M˜R and g˜max. The median M˜Amed of 100 different M˜A is plotted in Figure 3 as a function of M˜R and g˜max. Parameters M˜R and M˜Amed are normalized by the number of all samples in the scheme N, which is equal to 1000 in this particular case. Other values of N lead to a very similar relationship between M˜R, g˜max, and M˜Amed. The results of studying the parameter space of the two-step sampling approach are shown in Figure 3. At the bottom-left corner of the plot, we observe that a maximum of 30% of random samples may be taken to fill in all gaps larger than three grid steps and not to exceed 50% of active samples. As the gap threshold increases, more random samples can be taken. Finally, with the gap threshold of 8 grid steps, the second step may not be needed because a random scheme with 50% of active samples will likely have the largest gap size of 8 grid steps. The relationship between three key parameters of two-step sampling: the number of random samples MR taken during the first step, the maximum gap size gmax achieved with the designated samples, and the number of active samples MA in the two-step scheme. This empirical relationship shows that, as the threshold on the maximum gap size increases, more random samples are present in the sampling scheme for a fixed number of active samples. During the second step, the random sampling scheme is scanned to identify positions of the gaps larger than a given threshold. Furthermore, each gap is filled with a minimum number of designated samples until the gap threshold is achieved. We perform the gap-filling process using two options: recursive and jittered. The recursive gap-filling option includes the following steps: (i)Generate a local random scheme for each identified gap; Generate a local random scheme for each identified gap; (ii)Merge the local random scheme into the scheme from step (i); Merge the local random scheme into the scheme from step (i); (iii)Compute the gaps in the updated scheme; Compute the gaps in the updated scheme; (iv)Repeat steps (i)–(iii) until all large gaps are filled. Repeat steps (i)–(iii) until all large gaps are filled. The neighboring samples outside of the selected gaps can be considered to avoid repetitive patterns while filling these gaps in. The jittered gap-filling option uses jittered sampling to fill in the large gaps. Other sampling techniques may be considered during the second step since the NSP does not impose any restrictions on the choice of designated samples. 3.3. Fourier Spectrum of Sampling Schemes Once the sampling scheme is created, its characteristics need to be verified before moving to large-scale experiments. The sampling scheme represents a sequence of ones and zeros. The analysis of such binary sequence cannot be straightforward, especially if the sequence is long, for instance, 1000 samples. For this purpose, we consider the average Fourier spectrum of sampling schemes fave defined as follows: where fj is the Fourier spectrum of the j-th sampling scheme, and j indexes the realizations of sampling schemes generated by the same method. In the overview of different sampling methods, we noted that there is no consensus about using coherence, which is directly related to the Fourier spectrum of the sampling scheme, as a measure of goodness of the sampling scheme. In this paper, we use the average Fourier spectrum to analyze the statistical properties of these schemes in the Fourier domain. The average Fourier spectrum can also indicate the presence of specific redundant sampling patterns in the sampling schemes. Figure 4 shows the Fourier spectra computed based on Equation (3) with J=1000 for each sampling method. We consider that each scheme has 180 active samples (MA=180) and 180 omitted samples (MO=180), and the total number of samples N equals 360. We normalized the average Fourier spectrum by the number of active samples. The Fourier spectra analysis for different sampling schemes: (a) random, (b) two-step with recursive gap-filling option and the maximum gap of 5, (c) two-step with recursive gap-filling option and the maximum gap of 4, (d) two-step with recursive gap-filling option and the maximum gap of 3, (e) two-step with jittered gap-filling option and the maximum gap of 3, and (f) jittered with maximum gap of 3. At the zero wavenumber, the value of the spectrum is equal to one. The sampling schemes generated by different methods contain different sampling patterns that are revealed by the average Fourier spectrum. For the random schemes, the values of the average Fourier spectrum do not change significantly for all non-zero wavenumbers (Figure 4a). In the case of two-step schemes with the maximum gap size of 5 grid steps, about 95% of active samples are random. Consequently, the average spectrum of two-step schemes with gmax=5 resembles the almost flat average spectrum of the random schemes (Figure 4b). As the number of random samples and the maximum gap size in two-step schemes decreases, the average spectrum changes its shape. For the case of two-step schemes with the maximum gap size of 3 grid steps, the average spectrum achieves its maximum at the normalized wavenumber of about 0.3. We observe that the gap-filling option may also influence the shape of the average spectrum. For instance, the average spectra for two-step schemes are different for recursive and jittered gap-filling options, even though the maximum gap size is the same in both cases (Figure 4d,e). Finally, for the jittered schemes, the values of the average spectrum linearly increase as the normalized wavenumber increases (Figure 4f). We also report the maximum value of the Fourier spectrum, which was computed as where fj is the Fourier spectrum of the j-th sampling scheme, and j indexes the realizations of sampling schemes generated by the same method. As it is shown in Figure 4, the curves corresponding to the maximum follow the shape of the average Fourier spectra. In the generated two-step schemes, the second-step samples are used to meet the maximum gap size requirement. And even though both gap-filling options, recursive and jittered, rely on random sampling, the second-step samples are placed not irrespectively to the surrounding first-step samples. This characteristic differentiates two-step sampling from uniform random sampling, where all samples are placed uniformly at random. 4. Numerical Experiments 4.1. Generating and Modifying Sampling Schemes In this section, we describe the sampling schemes further analyzed with the average Fourier spectrum, CS reconstruction tests, fold maps, and rose diagrams of an orthogonal survey. We define the initial sampling schemes as those generated by a sampling method without applying additional subsequent changes. Three types of initial sampling schemes are considered. One of them is generated by the jittered sampling method. The two-step approach with recursive and jittered gap-filling options generates two other types. Each scheme has 360 samples (N=360), and 180 out of 360 samples are active (MA=180). Thus, other 180 samples are omitted (MO=N−MA=180), which implies a decimation ratio of 50%. The maximum gap size in these schemes equals 3 grid steps. In our experiments, we are also interested in evaluating the performance of different sampling schemes whose several initial sample locations have been changed. We define the modified sampling schemes as those generated by a sampling method and subsequently modified by a specific algorithm. Studying the modified sampling schemes is particularly important for seismic CS applications because unforeseen physical obstructions in the field may lead to changes in the prepared survey layout. Also, some sources (or receivers) can be classified as unacceptable for processing and removed from the recorded data, introducing additional distortions to the initial acquisition layout. We constructed the following modification process to imitate distortions that may happen with a sampling scheme during its field deployment: (i)Identify vacant positions in an initial scheme; Identify vacant positions in an initial scheme; (ii)Modify the initial scheme by moving K samples from their old positions to new positions from step (i); Modify the initial scheme by moving K samples from their old positions to new positions from step (i); (iii)Repeat the modification process Ntrials times for each K and each scheme to consider different scenarios of new positions. Repeat the modification process Ntrials times for each K and each scheme to consider different scenarios of new positions. In this modification algorithm, the new positions of K samples are chosen from the vacant positions such that the maximum gap size in the modified scheme remains the same as in the initial scheme. Also, this procedure avoids creating large gaps because it is known that sampling schemes with large gaps are not desirable for seismic surveys. As the samples are reallocated, the old positions of K samples cannot be considered as the new locations for any reallocated sample. This setting intends to imitate obstacles that may block the prepared positions of the sources (receivers) in the field. Every input scheme will have Ntrials modified versions, and each of them can be stored for further analysis. It is also possible to choose one modified version among all modification trials Ntrials using a criterion for the sampling scheme characterization. The modified sampling schemes for our numerical tests have been generated for each of the three considered scheme types: jittered sampling, two-step sampling with recursive gap-filling option, and two-step sampling with jittered gap-filling option. The parameters of modified schemes include the number of all samples N, the number of active samples MA, the maximum gap size, and the number of modified samples K0 and the operator inf takes the infimum of a function. The signal x is compressible when the error of its best s-term approximation decays rapidly in s [12]. The seismic wavefield is not inherently sparse but can be transformed into a compressible signal. Namely, the wavefield w∈RN and its compressible version x∈CL are connected by the sparse transform matrix C∈CL×N where H is the Hermitian transpose. The transform matrix C can be an orthonormal basis (e.g., the Fourier transform, N=L) or a tight frame (e.g., the curvelet transform, 7N≈L). The choice of the transform depends on the complexity of the wavefield and the sorting of the seismic data volume. A detailed review of other useful sparse transforms with localized basis functions (e.g., seislet, shearlet, coutrolets, etc.) and evaluation of two-step sampling performance using these sparse transforms are out of the scope of this paper. Another essential element of the CS framework is the sensing matrix A∈CM×L (M0 and c2>0 are constants that depend on the properties of the sensing matrix A. A sensing matrix with small c1 and c2 may not necessarily promote the reconstruction of the signal z closest to the unknown signal x compared to other signals reconstructed from the measurements made by means of other sensing matrices. On the one hand, small c1 and c2 lead to a small theoretical reconstruction error (the right-hand side of Equation (A7)), and the corresponding sensing matrix may be favorable for measurements. On the other hand, the right-hand side of Equation (A7) provides only an upper bound on the empirical reconstruction error (the left-hand side of Equation (A7)), which means that this error may not be strictly controlled by the constants c1 and c2. The empirical reconstruction error may be controlled and predicted when a signal model is used to build a sensing matrix, e.g., in the case of time-lapse seismic. Appendix B. Tools for Analyzing CS-Based Sampling Schemes The sensing matrix A should be specially constructed to meet the requirements of the seismic survey and enable data reconstruction for omitted sources and receivers. The CS framework uses metrics and matrix properties to analyze the goodness of the sensing matrix for the data reconstruction process. For example, the ℓ2 robust null space property (NSP) may be used to verify whether the sensing matrix is suitable for the reconstruction of approximately sparse signals from noisy measurements. The authors of [12] define this property as follows: the matrix A∈CM×L is said to satisfy the ℓ2-robust NSP of order s (with respect to ∥·∥) with constants 0<ρ<1 and τ>0 if, for any set S⊂[L] with card(S)≤s, The ℓ2 robust NSP is not computationally tractable and thus not very applicable for CS-based scheme design. Another property, the restricted isometry property (RIP), is also computationally intractable. However, different classes of randomly constructed matrices satisfy the RIP with high probability. According to [12], a matrix A∈CM×L is said to satisfy the RIP of order s if there exists 0<δ<1 such that for all s-sparse vectors x∈CL. The smallest δ that satisfies the above inequality is called the restricted isometry constant (RIC) and is denoted as δs=δs(A). A more popular metric to assess the goodness of the CS-based scheme is the coherence due to its computability. According to [12], the coherence μ=μ(A) of the matrix A with columns a1,a2,…aL is defined as The coherence μ is bounded from above by one and from below by the Welch bound μW as follows: In most cases, the coherence is computed when the Fourier basis is considered for the sparse transform C. In this scenario, the coherence is equal to the maximum non-DC amplitude of the Fourier spectrum of the sampling scheme. From the CS theoretical point of view, to reconstruct the s-sparse signal, it is sufficient that the coherence satisfies the following inequality: In practice, this inequality is rarely satisfied. For example, suppose the coherence value of a seismic CS-based scheme is 0.1. Then, the seismic volume in the transform domain should have only 5 non-zero coefficients, which is a very unrealistic scenario. The Fourier spectrum of a sampling scheme is yet another tool to analyze the properties of CS-based schemes. Although the CS theoretical guarantees are available only for the maximum non-DC amplitude of the Fourier spectrum of sampling schemes, in practice, the analysis of such spectrum is commonly used in CS-based scheme design [18]. In a seismic survey, sources are coupled with receivers because each active receiver records the seismic wavefield generated by each active source. Furthermore, a unique source–receiver pair produces a trace in the recorded data volume. The distribution of traces within offsets, azimuths, and bins can be analyzed by means of attributes from seismic survey design. For example, a fold map splits a survey area into bins, showing how many traces belong to the same bin. Another example is a rose diagram, which displays the number of traces for each offset–azimuth bin. A detailed description of how to compute the fold maps, rose diagrams, and other attributes can be found in books on seismic acquisition design [49]. Appendix C. The ℓ2-Robust NSP and Additional Rows in the Sensing Matrix Let us build a new matrix A^∈CMA×L by appending new rows into the matrix A∈CMR×L: where the matrix B∈CMD×L includes the new additional rows and MR+MD=MA0. In other words, the following inequality holds: for any set S⊂[L] with card(S)≤s and v∈CL. Combining Equations (A14) and (A16), we deduce that Thereby, the new matrix A^∈CMA×L has the ℓ2-robust NSP of order s with constants 0<ρ<1 and τ>0 for all v∈CL and any set S⊂[L] with cardinality s. Author Contributions Conceptualization, A.T., M.B.W. and A.C.T.; data curation, A.T. and A.C.T.; formal analysis, A.T. and M.B.W.; funding acquisition, A.C.T.; investigation, A.T.; methodology, A.T., M.B.W. and A.C.T.; project administration, M.B.W. and A.C.T.; resources, M.B.W. and A.C.T.; software, A.T. and M.B.W.; supervision, M.B.W. and A.C.T.; validation, A.T.; visualization, A.T.; writing—original draft preparation, A.T.; writing—review and editing, A.T., M.B.W. and A.C.T. All authors have read and agreed to the published version of the manuscript. Institutional Review Board Statement Not applicable. Informed Consent Statement Not applicable. Data Availability Statement The seismic shot gather used in the CS reconstruction experiments is from the SEAM Barrett dataset. The latter can be purchased from the SEG Advanced Modeling Corporation (SEAM) website https://seg.org/seam/data-sets/. Conflicts of Interest The authors declare no conflict of interest. Funding Statement This research was funded by the Reservoir Characterization Project (RCP) consortium at the Colorado School of Mines. Footnotes Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. References Associated Data Data Availability Statement The seismic shot gather used in the CS reconstruction experiments is from the SEAM Barrett dataset. The latter can be purchased from the SEG Advanced Modeling Corporation (SEAM) website https://seg.org/seam/data-sets/.