- Library Home /
- Search Collections /
- Open Collections /
- Browse Collections /
- UBC Theses and Dissertations /
- Curvelet-based migration amplitude recovery
Open Collections
UBC Theses and Dissertations
Featured Collection
UBC Theses and Dissertations
Curvelet-based migration amplitude recovery 2010
pdf
Page Metadata
Item Metadata
Title | Curvelet-based migration amplitude recovery |
Creator |
P. Moghaddam, Peyman |
Publisher | University of British Columbia |
Date Created | 2010-05-04 |
Date Issued | 2010-05-04 |
Date | 2010 |
Description | Migration can accurately locate reflectors in the earth but in most cases fails to correctly resolve their amplitude. This might lead to mis-interpretation of the nature of reflector. In this thesis, I introduced a method to accurately recover the amplitude of the seismic reflector. This method relies on a new transform-based recovery that exploits the expression of seismic images by the recently developed curvelet transform. The elements of this transform, called curvelets, are multi-dimensional, multi-scale, and multi-directional. They also remain approximately invariant under the imaging operator. I exploit these properties of the curvelets to introduce a method called Curvelet Match Filtering (CMF) for recovering the seismic amplitude in presence of noise in both migrated image and data. I detail the method and illustrate its performance on synthetic dataset. I also extend CMF formulation to other geophysical applications and present results on multiple removal. In addition of that, I investigate preconditioning of the migration which results to rapid convergence rate of the iterative method using migration. |
Genre |
Thesis/Dissertation |
Type |
Text |
Language | Eng |
Collection |
Electronic Theses and Dissertations (ETDs) 2008+ |
Date Available | 2010-05-04 |
Rights | Attribution-NonCommercial 2.5 Canada |
DOI | 10.14288/1.0052987 |
Degree |
Doctor of Philosophy - PhD |
Program |
Geophysics |
Affiliation |
Science, Faculty of |
Degree Grantor | University of British Columbia |
Graduation Date | 2010-11 |
Campus |
UBCV |
Scholarly Level | Graduate |
Rights URI | http://creativecommons.org/licenses/by-nc/3.0/ |
URI | http://hdl.handle.net/2429/24421 |
Aggregated Source Repository | DSpace |
Digital Resource Original Record | https://open.library.ubc.ca/collections/24/items/1.0052987/source |
Download
- Media
- ubc_2010_fall_poormoghaddam_peyman.pdf
- ubc_2010_fall_poormoghaddam_peyman.pdf [ 7.94MB ]
- ubc_2010_fall_poormoghaddam_peyman.pdf
- Metadata
- JSON: 1.0052987.json
- JSON-LD: 1.0052987+ld.json
- RDF/XML (Pretty): 1.0052987.xml
- RDF/JSON: 1.0052987+rdf.json
- Turtle: 1.0052987+rdf-turtle.txt
- N-Triples: 1.0052987+rdf-ntriples.txt
- Citation
- 1.0052987.ris
Full Text
Curvelet-Based Migration Amplitude Recovery by Peyman P. Moghaddam B.Sc. in Electrical Engineering, Amirkabir University of Technology, Tehran, Iran, 1995 M.Sc. in Electrical Engineering, Amirkabir University of Technology, Tehran, Iran, 1998 A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY in The Faculty of Graduate Studies (Geophysics) THE UNIVERSITY OF BRITISH COLUMBIA (Vancouver) April 2010 c© Peyman P. Moghaddam 2010 Abstract Migration can accurately locate reflectors in the earth but in most cases fails to correctly resolve their amplitude. This might lead to mis-interpretation of the nature of reflector. In this thesis, I introduced a method to accurately recover the ampli- tude of the seismic reflector. This method relies on a new transform-based recovery that exploits the expression of seismic images by the recently devel- oped curvelet transform. The elements of this transform, called curvelets, are multi-dimensional, multi-scale, and multi-directional. They also remain approximately invariant under the imaging operator. I exploit these properties of the curvelets to introduce a method called Curvelet Match Filtering (CMF) for recovering the seismic amplitude in presence of noise in both migrated image and data. I detail the method and illustrate its performance on synthetic dataset. I also extend CMF formulation to other geophysical applications and present results on multiple removal. In addition of that, I investigate preconditioning of the migration which results to rapid convergence rate of the iterative method using migration. ii Table of Contents Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii Table of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xx Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . xxi Dedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxii Statement of Co-Authorship . . . . . . . . . . . . . . . . . . . . . xxiii 1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.1 Theme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 1.2 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 1.3 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 iii Table of Contents 2 Sparsity and continuity promoting seismic image recovery with curvelet frames . . . . . . . . . . . . . . . . . . . . . . . . 9 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 2.1.1 Existing approaches . . . . . . . . . . . . . . . . . . . 11 2.1.2 Our approach . . . . . . . . . . . . . . . . . . . . . . 12 2.1.3 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . 15 2.2 Approximation of the normal operator . . . . . . . . . . . . 18 2.2.1 Zero-order imaging operators . . . . . . . . . . . . . . 19 2.2.2 Curvelet frames . . . . . . . . . . . . . . . . . . . . . 20 2.2.3 Diagonal approximation of PsDO ’s . . . . . . . . . . 22 2.2.4 Decomposition of the normal operator . . . . . . . . 25 2.2.5 Estimation of the diagonal . . . . . . . . . . . . . . . 26 2.3 Stable seismic image recovery . . . . . . . . . . . . . . . . . . 31 2.3.1 Curvelet frames for seismic images . . . . . . . . . . . 31 2.3.2 Stable recovery . . . . . . . . . . . . . . . . . . . . . 35 2.3.3 Solution by iterative thresholding . . . . . . . . . . . 37 2.3.4 Stable seismic recovery . . . . . . . . . . . . . . . . . 38 2.4 Sparsity- and continuity-enhanced seismic image recovery . . 40 2.5 Practical considerations . . . . . . . . . . . . . . . . . . . . . 44 2.6 Numerical experiments . . . . . . . . . . . . . . . . . . . . . 45 2.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 2.7.1 Initial findings . . . . . . . . . . . . . . . . . . . . . . 56 2.7.2 Extension to prestack imaging . . . . . . . . . . . . . 59 2.7.3 Recent related work . . . . . . . . . . . . . . . . . . . 60 2.7.4 Choice of the reference vector . . . . . . . . . . . . . 60 iv Table of Contents Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62 3 Curvelet-based migration preconditioning and scaling . . 67 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 67 3.2 Problem formulation . . . . . . . . . . . . . . . . . . . . . . . 69 3.3 Preconditioning . . . . . . . . . . . . . . . . . . . . . . . . . 72 3.3.1 Left preconditioning by fractional differentiation . . . 73 3.3.2 Right preconditioning by scaling in the physical do- main . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 3.3.3 Right preconditioning by scaling in the curvelet do- main . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 3.4 Convergence of least-squares migration . . . . . . . . . . . . 76 3.5 Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 3.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80 3.7 Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . 81 Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86 4 Curvelet-based seismic data processing . . . . . . . . . . . . 89 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 89 4.2 Curvelets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 4.3 Common problem formulation by sparsity-promoting inver- sion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91 4.4 Seismic data recovery . . . . . . . . . . . . . . . . . . . . . . 93 4.4.1 Curvelet-based recovery . . . . . . . . . . . . . . . . . 93 4.4.2 Focused recovery . . . . . . . . . . . . . . . . . . . . . 94 v Table of Contents 4.5 Seismic signal separation . . . . . . . . . . . . . . . . . . . . 96 4.6 Migration-amplitude recovery . . . . . . . . . . . . . . . . . 99 4.7 Discussion and conclusions . . . . . . . . . . . . . . . . . . . 100 4.8 Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . 102 Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104 5 True amplitude depth migration using curvelets . . . . . . 108 5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 108 5.2 Outline of the chapter . . . . . . . . . . . . . . . . . . . . . . 110 5.3 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . 110 5.4 Diagonal Approximation of the Normal Operator . . . . . . 114 5.5 Curvelets and their invariance under normal operator . . . . 115 5.6 Diagonal approximation . . . . . . . . . . . . . . . . . . . . . 117 5.7 Practical Workflow . . . . . . . . . . . . . . . . . . . . . . . 119 5.8 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 5.8.1 Noise-free case . . . . . . . . . . . . . . . . . . . . . . 121 5.8.2 Noisy data . . . . . . . . . . . . . . . . . . . . . . . . 130 5.9 Tolerance analysis . . . . . . . . . . . . . . . . . . . . . . . . 136 5.10 Verification of approximation . . . . . . . . . . . . . . . . . . 138 5.11 Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . 138 5.12 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140 5.13 Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . 140 Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142 vi Table of Contents 6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146 6.1 Main contributions . . . . . . . . . . . . . . . . . . . . . . . 146 6.1.1 Sparsity and continuity promoting seismic image re- covery with curvelet frames . . . . . . . . . . . . . . . 147 6.1.2 Curvelet-based migration preconditioning and scaling 148 6.1.3 Curvelet-based seismic data processing . . . . . . . . 149 6.1.4 True amplitude depth migration using curvelets . . . 149 6.2 Follow-up work . . . . . . . . . . . . . . . . . . . . . . . . . . 150 6.2.1 Full waveform inversion . . . . . . . . . . . . . . . . . 150 6.2.2 3D true amplitude migration . . . . . . . . . . . . . . 150 6.3 Adding more priori knowledge to solution . . . . . . . . . . . 150 6.4 Current limitations . . . . . . . . . . . . . . . . . . . . . . . 151 6.4.1 Curvelet code . . . . . . . . . . . . . . . . . . . . . . 151 Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153 Appendices A Proofs of lemma 1 and theorem 1 . . . . . . . . . . . . . . . 154 B Reverse time wave equation migration . . . . . . . . . . . . 158 B.1 Single scattering . . . . . . . . . . . . . . . . . . . . . . . . . 158 B.2 Shot-geophone modeling and migration . . . . . . . . . . . . 159 Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161 vii List of Tables 2.1 Algorithm for the estimation of the diagonal via regularized least-squares inversion. The Lagrange multiplier, η, is in- creased up to the point that all entries in the vector for the diagonal are positive. . . . . . . . . . . . . . . . . . . . . . . . 30 2.2 Sparsity-and continuity-enhancing recovery of seismic ampli- tudes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 viii List of Figures 2.1 Example of the recovery from different percentages of curvelet coefficients. The data set concerns a migrated image derived from the Mobil data set, (a) near perfect recovery from p = 99% of the coefficients, (b) recovery from only p = 3% of the coefficients, (c) the difference between near perfect recovery (a) and approximate recovery (b). This difference does not contain substantial coherent energy. . . . . . . . . . . . . . . 16 ix List of Figures 2.2 Invariance of curvelets under the discretized normal operator Ψ for a smoothly varying background model (a so-called lens model see Fig. 2.4(a)). Three coarse-scale curvelets in the physical domain before (a) and after application of the nor- mal operator (b) in the physical (a-b) and Fourier domain (e-f). The results for three fine-scale curvelets are plotted in (c-d) for the physical domain and in (g-h) for the Fourier domain. Remark: The curvelets remain close to invariant under the normal operator, a statement which becomes more accurate for finer scale which is consistent with Theorem 1. The example also shows that this statement only holds for curvelets that are in the support of the imaging operator ex- cluding steeply dipping curvelets. . . . . . . . . . . . . . . . . 17 2.3 Spatial and frequency representation of curvelets, (a) four dif- ferent curvelets in the spatial domain at three different scales, (b) dyadic partitioning in the frequency domain, where each wedge corresponds to the frequency support of a curvelet in the spatial domain. This figure illustrates the microlocal cor- respondence between curvelets in the physical and Fourier do- main. Curvelets are characterized by rapid decay in the physi- cal space and of compact support in the Fourier space. Notice the correspondence between the orientation of curvelets in the two domains. The 90◦ rotation is a property of the Fourier transform. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 x List of Figures 2.4 Stylized example of the diagonal estimation for a smooth ve- locity model and a reflectivity consisting of three reflection events including a fault, (a) the smooth background model with the low-velocity lens. This velocity model is used to cal- culate the linearized scattering and migration operators, (b) the bandwidth limited reflectivity, (c) the imaged reflectiv- ity from data consisting of 32 shot records with 500 traces each, (d) the approximated normal operator for diag(ũ) es- timated with η = 1. The bandwidth limited reflectivity in (b) and the image in (c) served as input for the reference and ’data’ vectors for the diagonal estimation procedure out- lined in Table 2.1. Remarks: the main dimming of the normal operator is captured by the diagonal approximation. The rel- ative `2-error between the actual and the approximate normal operator is 6.1%. . . . . . . . . . . . . . . . . . . . . . . . . . 32 2.5 Estimates for the diagonal ũ are plotted in (a-d) for increas- ing η = {0.01, 0.1, 1, 10}. The diagonal is estimated accord- ing the procedure outlined in Table 2.1 with the reference and ’data’ vectors, v and b, plotted in Fig. 2.4(b) and 2.4(c). As expected the diagonal becomes more positive for increasing η. 33 xi List of Figures 2.6 Decay rate for the curvelet coefficients compared to the de- cay for the coefficients of Fourier (dashed line) and discrete wavelet transforms (dot-dashed line), (a) decay rate for the sorted curvelet coefficients (solid line) of the real migrated im- age included in Fig. 2.1(a), (b) the same as (a) but now for the synthetic reflectivity of the SEG/EAGE AA′ salt model (cf. Fig. 2.8(a)). The decay for the curvelet transform clearly compares favorably to these other two transforms. . . . . . . 35 2.7 The SEG/EAGE AA′ salt model (Aminzadeh et al., 1997), (a) The original velocity model, (b) the smoothed velocity model with a window size of 720× 720m. Remark: Imaging this model is a challenge because of the presence of the high- velocity (bright) salt. . . . . . . . . . . . . . . . . . . . . . . . 47 2.8 Imaging of the SEG/EAGE AA′ salt model (Aminzadeh et al., 1997), (a) reflectivity defined in terms of the band-pass fil- tered difference between the smoothed velocity model and a slightly less smoothed velocity model, (b) the imaged reflec- tivity according to Eq. 2.24. The main reflection events are present but suffer from deteriorated amplitudes, especially under the high-velocity salt and for steep reflectors and faults. 48 xii List of Figures 2.9 Images that are input to the diagonal approximation scheme outlined in Table 2.1, (a) the reference vector, r, derived from Fig. 2.8(b), after applying corrections for the spherical- divergence and receiver-array taper, (b) the demigrated-migrated reference vector, b = Ψr, that serves as ’data’ for the diag- onal estimation, (c) diagonal approximation on the original bandwidth-limited reflectivity plotted in (a). This diagonal approximation byAATm captures the normal operator,Ψm, quite well. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 2.10 Gradient vectors for the reference vector r plotted in Fig. 2.9(a). The gradients (white ↑’s) are used to calculate the tangential directions along which the additional anisotropic smoothing is applied. . . . . . . . . . . . . . . . . . . . . . . . 51 2.11 Images after nonlinear recovery (P). (a) result with the `1- norm recovery only (β = 0); (b) recovery with the combined `1 and continuity recovery for α = β = 1/2. The amplitudes are recovered. The anisotropic diffusion successfully removes the artifacts. . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 xiii List of Figures 2.12 Recovery of the amplitudes after normalization with respect to the first event, (a) seismic traces of the original reflectivity (dot-dashed line), the migrated reflectivity (dashed line) and the recovered reflectivity (solid line), (b) amplitudes along the bottom most horizontal reflector, (c) average amplitude spectra of the depth-dependent reflectivity. The original and recovered spectra are normalized to match. Remarks: The nonlinear recovery corrects for the amplitudes and restores the spectrum. . . . . . . . . . . . . . . . . . . . . . . . . . . . 54 2.13 Image amplitude recovery for noisy data (SNR 3dB), (a) noise image according to Eq. 2.24, (b) image after nonlin- ear recovery from noisy data (P). The clearly visible non- stationary noise in (a) is removed during the recovery while the amplitudes are also restored. Steeply dipping reflectors denoted by the arrows are also well recovered. . . . . . . . . 57 2.14 ’Denoising’ of a shot record, (a) the noise-free data received by the receiver array, (b) noisy data with a SNR of 3 dB, (c) forward modeled data after amplitude recovery. Observe the significant improvement in the data quality, reflected in an increase for the SNR of 19.2 dB. . . . . . . . . . . . . . . . . 58 xiv List of Figures 3.1 The SEG/EAGE AA′ salt model, (a) reflectivity defined by the high-pass filtered velocity model, (b) smoothed velocity model, (c) the migrated image according to Equation 3.1. This image suffers from deteriorated amplitudes, especially under the high-velocity salt and for steep reflectors and faults. 82 3.2 Migrated images for different levels of preconditioning, (a) result for left preconditioning (level I, cf. Equation 3.6), (b) result for left-right (including depth-correction) precondition- ing (level II, cf. Equation 3.7), (c) the same but now including curvelet-domain scaling (level III, cf. Equation 3.10). . . . . . 83 3.3 Residual decays for different levels of preconditioning. The dotted blue lines corresponds to least-squares migration with- out preconditioning, the dash-dotted lines to level I precon- ditioning, the dashed black lines to level II preconditioning, and the red solid lines to level III preconditioning. This is offset by one iteration to account for the overhead, (a) plot for the decay of the data-space normalized residues µk as a function of the number of LSQR iterations, (b) the same but now for the model-space normalized residuals νk. . . . . . . . 84 3.4 Least-squares migration without and with preconditioning juxtaposed with the original reflectivity (in light blue), (a) least-squares migrated image, (b) least-squares image with level III preconditioning. Notice the improvement in the re- covered reflectivity. . . . . . . . . . . . . . . . . . . . . . . . . 85 xv List of Figures 4.1 Comparison between 3-D curvelet-based recovery by sparsity- promoting inversion with and without focusing, (a) fully sam- pled real SAGA data shot gather, (b) randomly subsampled shot gather from a 3-D data volume with 80% of the traces missing in the receiver and shot directions, (c) curvelet-based recovery, (d) curvelet-based recovery with focusing. Notice the improvement (denoted by the arrows) from the focusing with the primary operator. . . . . . . . . . . . . . . . . . . . 95 4.2 3-D Primary-multiple separation withP for the SAGA dataset, (a) near-offset section including multiples, (b) the SRME- predicted multiples, (c) the estimated primaries according to `2-matched filtering, (d) the estimated primaries obtained with P. Notice the improvement, in areas with small 3-D effects (ellipsoid) and residual multiples. . . . . . . . . . . . . 98 4.3 Image amplitude recovery for a migrated image calculated from noisy data (SNR 3dB), (a) image with clutter, (b) im- age after nonlinear recovery. The clearly visible non-stationary noise in (a) is mostly removed during the recovery while the amplitudes are also restored. Steeply dipping reflectors (de- noted by the arrows) under the salt are also well recovered. . 101 xvi List of Figures 5.1 Spatial and frequency representation of curvelets, (a) four dif- ferent curvelets in the spatial domain at three different scales, (b) dyadic partitioning in the frequency domain, where each wedge corresponds to the frequency support of a curvelet in the spatial domain. This figure illustrates the microlocal cor- respondence between curvelets in the physical and Fourier do- main. Curvelets are characterized by rapid decay in the phys- ical space and of compact support in the Fourier space. No- tice the correspondence between the orientations of curvelets in the two domains. The 90o rotation is a property of the Fourier transform. Courtesy of Herrmann et al. (2008). . . . 115 5.2 Conflicting dips velocity model and reflectivity (a) velocity model, (b) smooth velocity model used to generate our exam- ple, (c) reflectivity generated by subtracting smooth velocity model from original one . . . . . . . . . . . . . . . . . . . . . 122 5.3 Diagonal approximation of normal operator, (a) reference im- age (depth corrected migrated image), (b) normal operator on the reference image (i.e., Ψr) (c) approximate normal op- erator on the reference image (i.e., CTWCr), approximation error is %3.71 . . . . . . . . . . . . . . . . . . . . . . . . . . 124 xvii List of Figures 5.4 Curvelet-domain representation of the curvelet vector obtained from (a) reference vector, (b) scaling coefficients without smoothing constraints, (c), scaling coefficients with smooth- ing constraint. The different sub-images represent the curvelet coefficients at different scales (coarsest in the center) and dif- ferent angles. . . . . . . . . . . . . . . . . . . . . . . . . . . . 126 5.5 Examples of the solution for optimization problem QP with soft-thresholding, (a) λ = 0, (b) λ = σb (c) λ = 2σb . . . . 127 5.6 Examples of the solution for optimization problem BPDN with SPGL1, (a) = σb, (b) = 50σb (c) = 100σb . . . . 129 5.7 Reflectivity model, the vertical lines are locations of investi- gation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130 5.8 Amplitude analysis of the recovered images using proposed methods, (a) offset=3480 (m) , (b) offset=5160 (m), (c) off- set=5880 (m) . . . . . . . . . . . . . . . . . . . . . . . . . . . 131 5.9 Amplitude analysis of the recovered images using proposed methods, (a) offset= 8640 (m), (b) offset=10080 (m) , (c) offset=12720 (m), note the salt bottom amplitude mismatch 132 5.10 Synthetic data, (a) before , (b) after adding the noise, SNR = 10dB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133 5.11 Depth corrected migrated noisy image . . . . . . . . . . . . . 134 5.12 Examples of the solution for optimization problem QP with soft-thresholding, (a) λ = σb, (b) λ = 2σb (c) λ = 3σb . . . 135 5.13 Examples of the solution for optimization problem BPDN with SPGL1, (a) = 10σb, (b) = 140σb (c) = 200σb . . 137 xviii List of Figures 5.14 Comparison between the approximation terms in Equation 5.21, (a) left hand term El = CTW− 1 2CCTWx , (b) right hand term Er = CTW 1 2x, relative error is 4.09 percent. . . . 139 xix Preface This thesis was prepared with Madagascar, a reproducible research soft- ware package available at rsf.sf.net, in such a way that most of the repro- ducible results are linked to the code that generated them. Reproducibility facilitates the dissemination of knowledge not only within the Seismic Lab- oratory for Imaging and Modeling (SLIM) but also between SLIM and its sponsors, and more generally, the entire research community. The programs required to reproduce the reported results are Mada- gascar programs written in C/C++, MatlabR©, or Python. The numeri- cal algorithms and applications are mainly written in Python as part of SLIMpy (slim.eos.ubc.ca/SLIMpy) with a few exceptions written in MatlabR© or Python. xx Acknowledgments I am greatly indebted to my advisor Felix Herrmann. A few words of acknowledgment are just not enough to express my gratitude and appreci- ation for all he has done for me. I also wish to thank him for his kind and sometimes (very) energetic encouragements. I would like to express my gratitude to Reza Shahidi, Deli Wang, Cody Brown, Gilles Hennenfent, and the remainder of the SLIM team for their help and friendships. I was very fortunate to be part of this group. The next persons who have influenced my scientific life are Chris Stolk, William Symes, Uri Ascher and Chen Grief. They helped my research in- cursions in the areas of applied mathematics and optimization. I would like to thank them for their time and patience. My appreciation goes to Total E&P for the hospitality during my intern- ship. In particular, I wish to thank Eric Dussaud, and Paul Williamson for great technical discussions. I would like to express my gratitude to Chen Grief, Michael Bostock, Doug Oldenburg for serving on my supervisory committee. Finally, many thanks to my family for their love, support and encour- agement throughout my studies. xxi To my family xxii Statement of Co-Authorship Chapter 2 was published jointly with Felix J. Herrmann and Chris Stolk in the mathematical literature where alphabetical ordering of authors is customary. Aside from providing the numerical examples for this paper, I co-authored the text for the examples and I wrote the description of the part that involved our nonlinear solver. Chapter 3 was published with Felix J. Herrmann, Cody R. Brown, and Yogi A. Erlangga. My contribution consisted of the development of the curvelet matched filter and the design of the imaging experiment. Cody assisted with generating the results and Felix wrote most of the manuscript with significant inputs from me and the other co-authors. Chapter 4 is a review paper highlighting three distinctively different ap- plications of the curvelet transform to seismic data processing and imaging. I was responsible for the imaging example and I authored the corresponding text. Chapter 5 was prepared for submission for publication jointly with Fe- lix J. Herrmann with input from Reza Shahidi. I was responsible for all examples and I authored the text. xxiii Chapter 1 Introduction Exploration seismology is a complex and costly operation. On land, dynamite or Vibroseis sources can be used to send energy into the subsurface. This energy propagates and partially reflects at interfaces due to a change in rock properties. The reflected wavefield is recorded at the earth’s surface by an array of geophones. At sea, the principle remains the same but the seismic source is typically an air gun and the receivers are hydrophones on streamer lines towed by a seismic vessel. Seismic processing is a technique widely used by the oil industry to explore and identify potential oil-rich areas into the earth using the recorded data. Seismic migration is vital part of seismic processing. Migration generates an image of the subsurface that is finally interpreted by geo-scientists. Migration difficulties generally arise from assumptions made in algo- rithms, that are not met by underlying theory. In particular, modern mi- gration algorithms can accurately position reflecting features in the Earth’s subsurface, however they generally do not correctly resolve the amplitudes. This problem has different causes. First, the migration operator is the ad- joint but not the inverse of the linearized Born modeling or scattering op- erator (see e.g. Guitton, 2004; Claerbout, 1985; Gray, 1997; Symes, 2008). 1 1.1. Theme Second, There are also other sources contributing to this inaccuracy in- cluding geometrical spreading of wavefront energy, acquisition limitations, velocity heterogeneity and anomalous focusing. There exists a wide variety of migration amplitude recovery techniques: • iterative-based methods use migration and scattering operators and iteratively recover the amplitude of the reflectors through a Krylov subspace-based method such as LSQR or conjugate-gradient (Kuehl and Sacchi, 2003). They typically fairly computationally intensive • match filter-based methods represent another type of amplitude recovery approach that include the migrated and re-migrated image. They require ”only one” time modeling and migration of the image and computationally cheaper than iterative based methods. (Guitton, 2004; Claerbout, 1985; Rickett, 2003; Symes, 2008) are examples of this approach. 1.1 Theme The method proposed in this thesis follows the match filter-based ap- proach. The main theme of this thesis is a practical, robust, and geomet- rical—i.e., transform-based—approach to the seismic migration amplitude recovery problem. The motivation of this approach is that two key features of seismic data are, in my opinion, not used to their full extent in existing approaches, namely • Representation of the seismic normal operator in transform 2 1.1. Theme domain The seismic normal operator typically belongs to a class of operators called Pseudo-differential operators. These operators have interesting properties to make them suitable to represent in the trans- form domain. • Strong geometrical structure Seismic images are a spatio-temporal sampling of the reflectors in the Earth’s subsurface. These reflectors are mathematically continuous functions of reduced dimension (i.e., curves in 2D or sheets in 3D) To make the most of these properties, my approach uses the curvelet transform (Candès and Donoho, 2004) which is data-independent, multi- scale, and multidirectional. The elements of this transform, the curvelets, are localized in the frequency domain and of rapid decay in the physical domain. They are very efficient at representing curve-like singularities— e.g., reflectors. In other words, only a few curvelets are needed to represent the complexity of real seismic image. I use this piece of information, called sparsity, to help solve the imaging problem. The idea of sparsity-promoting inversion is in itself not new to geophysics. The other interesting property of the curvelets is their behavior under the action of acoustic wave propagator (e.g., Candès and Demanet, 2003; Smith, 1998, 1997) . A theorem in chapter 2 proved that the curvelets are approximately invariant under the action of the normal operator . Further- more I adopt the result of this theorem to introduce match-filtering approach for estimation of this operator. 3 1.2. Objectives 1.2 Objectives The objectives of this thesis are two-fold: • develop an in-depth understanding of successful curvelet match filter- ing for the approximation of normal operator and its key ingredients. • formulate a practical sparsity-promoting seismic migration amplitude recovery algorithm. 1.3 Outline In chapter 2, I observe that curvelets, as a directional frame expansion, lead to sparsity of seismic images and exhibit invariance under the normal operator of the linearized imaging problem. Based on this observation I derive a method for stable recovery of the migration amplitudes from noisy data. The method corrects the amplitudes during a post-processing step after migration, such that the main additional cost is one application of the normal operator, i.e. a modeling followed by a migration. Asymptotically this normal operator belongs to a class of operators known as pseudodiffer- ential operators, for which a diagonal approximation in the curvelet domain is derived, including a bound for its error and a method for the estimation of the diagonal from a compound operator consisting of discrete implementa- tions for the scattering operator and its adjoint the migration operator. The solution is formulated as a nonlinear optimization problem where sparsity in the curvelet domain, as well as continuity along the imaged reflectors, are jointly promoted. To enhance sparsity, the `1-norm on the curvelet 4 1.3. Outline coefficients is minimized, while continuity is promoted by minimizing an anisotropic diffusion norm on the image. The performance of the recovery scheme is evaluated with a time-reversed ’wave-equation’ migration code on synthetic datasets, including the complex SEG/EAGE AA′ salt model. In chapter 3, I introduce a preconditioner for the inversion of the lin- earized Born scattering operator. This preconditioner approximately cor- rects for the “square root” of the normal operator. This approach consists of three parts, namely (i) a left preconditioner, defined by a fractional time integration designed to make the migration operator zero order, and two right preconditioners that apply (ii) a scaling in the physical domain ac- counting for a spherical spreading, and (iii) a curvelet-domain scaling that corrects for spatial and reflector-dip dependent amplitude errors. I show that a combination of these preconditioners lead to a significant improve- ment of the convergence for iterative least-squares solutions to the seismic imaging problem based on reverse-time migration operators. In chapter 4, I show that other geophysical problems—e.g., focused re- covery, seismic signal separation, and migration amplitude recovery—can be re-cast in the formulation used for curvelet sparsity inversion formulation. This puts in a broader perspective the insights gained during the develop- ment of curvelet sparsity inversion formulation. In chapter 5, I focused on the practical approach of amplitude recovery method. I revisit the amplitude recovery method and extend it threefold. First, I replace the linear least-squares formulation for the estimation of the curvelet-domain coefficients (Herrmann et al. (2008)) by a non-linear least-squares formulation that imposes positivity on the scaling coefficients. 5 1.3. Outline This comes from the fact that the normal operator is positive semi-definite. Second, I consider the noise term both in the data and that generated by migration due to incomplete data (i.e., few shots). Third, I investigate two methods for recovery and compare them in terms of quality and performance, a soft thresholding-correction technique and a sparsity promotion technique. In Chapter 6, I summarize the work done in this thesis, and discuss some of its aspects in a broader context. Conclusions and recommendations for future research follow. Appendix A contains further details about the curvelet transform. It proves the theorem and associated lemma regarding the invariance of the curvelet elements under the action of normal operator. This appendix pairs with chapter 2. 6 Bibliography Candès, E. J. and L. Demanet, 2003, Curvelets and Fourier integral opera- tors: Compte Rendus de l’Académie des Sciences, Paris, 336, no. 1, 395 – 398. Candès, E. J. and D. L. Donoho, 2004, New tight frames of curvelets and optimal representations of objects with C2 singularities: Communications on Pure and Applied Mathematics, 57, no. 2, 219 – 266. Claerbout, J., 1985, Earth soundings analysis, processing versus inversion: Blackwell Scientific Publications. Gray, S. H., 1997, True amplitude seismic migration: A comparison of three approaches: Geophysics, 62, 929–936. Guitton, A., 2004, Amplitude and kinematic corrections of migrated images for nonunitary imaging operators: Geophysics, 69, 1017–1024. Herrmann, F. J., P. P. Moghaddam, and C. Stolk, 2008, Sparsity- and continuity-promoting seismic image recovery with curvelet frames: Ap- plied and Computational Harmonic Analysis, 24, 150–173. Kuehl, H. and M. Sacchi, 2003, Least-squares wave-equation migration for AVP/AVA inversion: Geophysics, 68, 262–273. Rickett, J., 2003, Illumination based normalization for wave-equation depth migration: Geophysics, 68, 208–221. 7 Chapter 1. Bibliography Smith, H., 1997, A Hardy space for Fourier integral operaotrs: Geom. Anal., 7, 784–842. ——–, 1998, A parametrix construction for wave equations with C1,1 coef- ficients: , Ann. Inst. Fourier, Grenoble, 48, 797–835. Symes, W., 2008, Optimal scaling for reverse time migration: submitted to Geophysics. 8 Chapter 2 Sparsity and continuity promoting seismic image recovery with curvelet frames 2.1 Introduction This paper introduces a general-purpose algorithm for the stable recovery of the amplitudes in seismic images. The method involves the inversion in dimension d of a zero-order pseudodifferential operator (PsDO ) of the type ( Ψf ) (x) = ∫ Rd e−ix·ξa(x, ξ)f̂(ξ)dξ (2.1) with f̂(ξ) = 1 (2pi)d ∫ Rd f(x)eix·ξdx (2.2) A version of this chapter has been published. Herrman, F.J., Moghaddam, P.P. and Stolk, C. (2008) Sparsity- and continuity-promoting seismic image recovery with curvelet frames. Applied and Computational Harmonic Analysis, Vol. 24, No. 2, pp. 150-173, 2008. 9 2.1. Introduction the Fourier transform with x, ξ ∈ Rd the spatial coordinate and wavenum- ber vectors and a(x, ξ) the symbol of the pseudodifferential operator. Under certain conditions that include no grazing rays and high-frequency asymp- totics, the above linear operator describes seismic data after imaging (ten Kroode et al., 1998; de Hoop and Brandsberg-Dahl, 2000; Stolk and Symes, 2003). A seismic image is derived from noisy data given by d(xs, xr, t) = ( Km ) (xs, xr, t) + n(xs, xr, t) (2.3) with K the Born scattering operator, m(x) the model with the reflectivity and n zero-mean standard deviation σ white Gaussian noise. The seismic data volume is a function of the source and receiver positions, xs ∈ Rd−1 and xr ∈ Rd−1 and of time, t ∈ R+0 . After applying the migration operator KT , (the symbol T is reserved for the transpose), to the data volume an “image” is obtained according to (KTd)(x) = ( KTKm ) (x) + ( KT ) n(x) y(x) = ( Ψm ) (x) + e(x). with x ∈ Rd (2.4) This equation for the image y corresponds to the normal equation and con- tains contributions from the normal operator, Ψ, as well as from imaged noise e. The operator K preserves the singularities in y, i.e., KTKm ≈ Id m with Id the identity matrix and the symbol ≈ indicating ”approximated by” in the locations of the singularities. With the increased demand for high- quality images, the above approximation of Ψ ≈ Id is no longer justifiable 10 2.1. Introduction because it may lead to errors in the relative amplitudes of the imaged re- flectors. Our main goal is to derive a method that recovers these imaged amplitudes, i.e. estimate models m from “images” y with clutter. 2.1.1 Existing approaches With the increasing demand for hydrocarbons and the increasingly hard to image (subsalt), the recovery of correct seismic amplitudes has become more necessary. Approaches in the extensive literature on this topic range from least-squares migration, where the normal or Gramm matrix is inverted with Krylov subspace methods (see e.g. Nemeth. et al., 1999; Chavent and Plessix, 1999; de Hoop and Brandsberg-Dahl, 2000), to high-frequency (mi- crolocal) methods (see e.g. ten Kroode et al., 1998; de Hoop and Brandsberg- Dahl, 2000; Stolk and Symes, 2003), where the normal operator is inverted based on ray-asymptotic arguments, to methods that apply a diagonal scal- ing (see e.g. (Rickett, 2003; Guitton, 2004; Plessix and Mulder, 2004) and more recently Symes (2006a)) to approximately invert the Gramm operator of ’wave-equation’ migration. These scaling methods either calculate the weighting analytically (ten Kroode et al., 1998; Plessix and Mulder, 2004), based on certain assumptions regarding the acquisition and velocity model, or estimate the diagonal weighting from the action of the normal operator on some reference vector, an idea first suggested by Symes and reported in (Claerbout and Nichols, 1994). In this paper, a similar approach is fol- lowed, where the normal operator is replaced by a diagonal matrix acting on the curvelet coefficients of the image (Candes et al., 2006a; Hennenfent and Herrmann, 2006). 11 2.1. Introduction 2.1.2 Our approach The computational cost of evaluating the migration operator is O(nd+2) for a d-dimensional image with n samples in each direction. This large cost makes it a major challenge to conduct least-squares migration for d > 2 or large n Nemeth. et al. (1999); Chavent and Plessix (1999); de Hoop and Brandsberg-Dahl (2000); Hu et al. (2001); Kuhl and Sacchi (2003); Plessix and Mulder (2004); Hu et al. (2001). We address this issue by exploit- ing recently developed curvelet frames. These frame expansions compress seismic images (see e.g. Candes et al., 2006a; Hennenfent and Herrmann, 2006, for the compression of seismic data and Fig. 2.1 for the compression of a typical migrated seismic image) and consist of a collection of frame elements ’curvelets’ that are approximately invariant under PsDO ’s. These two properties allow for the development of an approach similar to the so- called wavelet-vaguelette method (WVD) (see e.g. Donoho, 1995; Candes and Donoho, 2000b; Lee and Lucier, 2001). In this approach, scale-invariant homogeneous operators are nonlinearly inverted, using the eigenfunction-like behavior of wavelets and curvelets. Our main contribution to earlier ideas on stable seismic image recovery (see e.g. Herrmann, 2003) is threefold: First, the WVD method is extended to expanding PsDO ’s with respect to redundant curvelet frames. Second, it is observed that order-zero pseudodifferential operators with homogeneous principal symbols act approximately as a multiplication on curvelets. This property implies that such operators can be approximated by a diagonal matrix acting on curvelet coefficients of an image, followed by an inverse 12 2.1. Introduction curvelet transform. We refer to this procedure as a curvelet-domain diag- onal approximation (or weighting) that becomes more accurate for smaller scales (higher frequencies). Third, a formulation for the amplitude recov- ery is presented in terms of a nonlinear sparsity-and continuity-enhancing optimization problem. After discretization, the nonlinear optimization problem for the seismic amplitude recovery (see e.g. Tsaig and Donoho, 2006; Candes et al., 2006b; Daubechies et al., 2005) has the following form P : x̃ = argminx J(x) = Js(x) + Jc(x) subject to ‖y −Ax‖2 ≤ m̃ = ( AT )† x̃. (2.5) During the optimization, the vector x is optimized with respect to the penalty functional J(x) and the `2-data misfit. The penalty functional J(x) is combination of sparsity penalty function Js(x) and continuity penalty function Jc(x). The term sparsity vector is used for x to stress the point that the magnitude-sorted coefficients of this vector are of rapid decay, by virtue of compression by curvelet frames. The above optimization problem is solved for the model by finding a coefficient vector x̃ that minimizes the penalty term subject to fitting the data to within a noise-level dependent tolerance level . The ’tilde’ symbol (˜) is reserved for vectors that solve a nonlinear optimization problem. The penalty functional J(x) is designed to promote the sparsity of the image in the curvelet domain (i.e., Js(x)) as well as continuity along the imaged reflectors (i.e., Jc(x)). Bold lowercase symbols refer to discretized vectors 13 2.1. Introduction and bold uppercase symbols to discretized operators. Non-boldface symbols refer to continuous infinite dimensional functions and operators. Estimates for the model m̃ are calculated by applying the pseudo-inverse (denoted by the symbol †) of the analysis matrix to x̃. This analysis matrix is given by the adjoint of the synthesis matrix (AT ). This synthesis is given by the diagonally-weighted curvelet synthesis matrix with a weighting designed such that AAT r ' Ψr. (2.6) In this expression, r represents an appropriately chosen discrete reference vector and Ψ the discrete normal operator, formed by compounding the discrete scattering and its transpose the migration operator, i.e.,Ψ := KTK with the symbol := denoting ’defined as’. The symbol ' is used to denote that this expression is approximate, a statement we will make more precise with a bound on the error. After forming the normal operator with one’s favorite numerical imple- mentation for the migration operator and its adjoint, our amplitude recovery method consists of the following steps: 1. Calculate an appropriate reference vector, r, from the data that is close enough to the unknown image. Typically, this reference vector is defined by a migrated image to which a simple amplitude correction is applied; 2. Estimate a diagonal curvelet-domain weighting that approximates the normal operator on the reference vector. This diagonal defines the synthesis matrix A; 14 2.1. Introduction 3. Estimate x̃ by solving the nonlinear optimization problem P. This program inverts the synthesis matrix; 4. Calculate the model m̃ from the recovered coefficient vector x̃ through the pseudo-inverse of AT . The proposed recovery method derives from two essential properties of curvelet frames, namely the ability of this signal representation to compress seismic images and the invariance of curvelets under the normal operator. Fig. 2.1 and 2.2 are included to stress the importance of these properties by confirming that a real migrated image can accurately be recovered from only 3% of the curvelet coefficients, while curvelets remain invariant under the normal operator defined for a smooth background velocity model. Dur- ing this paper, we use the first property to formulate a sparsity promoting seismic image in term of a nonlinear optimization and the second property for diagonal approximation of the normal operator in curvelet domain. 2.1.3 Outline First, a diagonal approximation of the normal operator in the curvelet do- main is derived. The error of this approximation is bounded, using proper- ties of the curvelet tiling of phase space. In this derivation, use is made of the property that the normal operator can be described as a zero-order pseu- dodifferential operator. Next, a method is proposed to estimate the diago- nal weighting matrix that uses the property that the symbol of the normal operator is smooth in phase space and positive. This diagonal approxima- tion leads to an approximate seismic image representation by a diagonally- 15 2.1. Introduction (a) (b) (c) Figure 2.1: Example of the recovery from different percentages of curvelet coefficients. The data set concerns a migrated image derived from the Mobil data set, (a) near perfect recovery from p = 99% of the coefficients, (b) recovery from only p = 3% of the coefficients, (c) the difference between near perfect recovery (a) and approximate recovery (b). This difference does not contain substantial coherent energy. 16 2.1. Introduction (a) (b) (c) (d) (e) (f) (g) (h) Figure 2.2: Invariance of curvelets under the discretized normal operator Ψ for a smoothly varying background model (a so-called lens model see Fig. 2.4(a)). Three coarse-scale curvelets in the physical domain before (a) and after application of the normal operator (b) in the physical (a- b) and Fourier domain (e-f). The results for three fine-scale curvelets are plotted in (c-d) for the physical domain and in (g-h) for the Fourier domain. Remark: The curvelets remain close to invariant under the normal operator, a statement which becomes more accurate for finer scale which is consistent with Theorem 1. The example also shows that this statement only holds for curvelets that are in the support of the imaging operator excluding steeply dipping curvelets. 17 2.2. Approximation of the normal operator weighted curvelet transform. With this representation, the seismic image amplitudes can be recovered by solving a nonlinear optimization problem. During this optimization problem, the image is recovered by minimizing the data mismatch while promoting sparsity in the curvelet domain and con- tinuity along imaged reflectors. This paper is concluded by applying the algorithm to a synthetic data set derived from the SEG AA′ salt model. 2.2 Approximation of the normal operator The primary focus of seismic imaging is to locate singularities in the earth’s elastic properties from seismic data recorded at the surface (Beylkin, 1984; de Hoop and Bleistein, 1997; ten Kroode et al., 1998; Stolk, 2000; Bleis- tein et al., 2001; Brandsberg-Dahl and de Hoop, 2003). A seismic survey consists of multiple seismic experiments in which both the location of the sources and receivers are varied along the surface. The acquired data is sub- sequently used to create images of the singularities in the subsurface. The main purpose of the recovery method presented in this paper is to recover the relative amplitudes of seismic images from data that is possibly contam- inated with noise. For this purpose, a diagonal approximation of the normal operator in the curvelet domain is presented. This approximation derives from the invariance properties of curvelets under the normal operator (see Fig. 2.2). The approximation leads to a SVD-like decomposition for the normal operator and makes the large-scale seismic image recovery problem amenable to a solution by nonlinear optimization. 18 2.2. Approximation of the normal operator 2.2.1 Zero-order imaging operators In the high-frequency limit, the scattering operator and the normal oper- ator can, under certain conditions on the medium and ray-geometry, be considered as Fourier Integral Operators (FIO’s) (ten Kroode et al., 1998). In dimension two (d = 2), the scattering operator, K, and its adjoint the migration operator, KT , can both be considered as FIO’s of order 14 , while the leading behavior for their composition, the normal operator Ψ, corre- sponds to an order-one invertible elliptic PsDO (ten Kroode et al., 1998; de Hoop and Brandsberg-Dahl, 2000; Stolk and Symes, 2003). To make this PsDOamenable to an approximation by curvelets, the following substitu- tions are made for the scattering operator and the model: K 7→ K (−∆)−1/2 and m 7→ (−∆)1/2m with ((−∆)αf)∧(ξ) = |ξ|2α · f̂(ξ), with ∆ is the Lapla- cian operator. Alternatively, these operators can be made zero-order by composing the data side with a 1/2-order fractional integration along the time coordinate, i.e., K 7→ ∂−1/2t K. After these substitutions, the normal operator Ψ becomes zero-order. Similar substitutions are made in the WVD methods where vaguelettes are introduced according to the same mappings. Our seismic image amplitude problem is different from the recovery of images in ill-posed problems such as inverting the Radon transform (Candes and Donoho, 2000b). In our case, the ill-posedness comes from small values of the symbol at certain positions and wave numbers and is not related to the ill-posedness stemming from the inversion of a normal operator that acts as a (fractional) inverse Laplacian. Our normal operator, without the above substitutions, acts as a (fractional) Laplacian. This behavior makes 19 2.2. Approximation of the normal operator the problem ill-posed for the low-frequencies. Since seismic data and the approximations of the operators are both high-frequency, this ill-posedness is negated. The ill-posedness that is a concern are the entries in the symbol of the normal operator that correspond to regions in the model space that are badly insonified. 2.2.2 Curvelet frames Curvelets are directional frames that represent a specific tiling of the two/three- dimensional frequency plane into multiscale and multi-angular wedges (see Fig. 5.1). Because the directional sampling increases every-other scale dou- bling, curvelets become more and more anisotropic at finer and finer scales. They become ’needle-like’ as illustrated in Fig. 5.1. Curvelets are localized in Fourier space and their smoothness in this domain leads to a rapid decay in the physical domain. Their effective support is given by ellipsoids with a width ∝ 2j/2 and length ∝ 2j and an angle θjl = 2pil2bj/2c with j the scale and l the angular index. Curvelets are indexed by the multi-index µ := (j, l, k) ∈ M with M the multi-index set running over all scales, j, angles, l, and positions k (see for details Candes et al., 2006a; Ying et al., 2005). Following Candes et al. (2006a), define the continuous curvelet family for x ∈ R2 as ϕµ(x) = 23j/4ϕ(DjRθjlx− k), (2.7) where • ϕ(x) is a smooth bell-shaped function in the horizontal direction and oscillatory in the vertical direction; 20 2.2. Approximation of the normal operator • Dj = diag(2j , 2bj/2c) is the parabolic scaling matrix; • Rθjl is the rotation matrix over angles θjl = 2pil2−bj/2c with 0 ≤ θjl < 2pi; • k = (k1, k2) ∈ Z2 the translation parameters. In the frequency domain, curvelets are compactly supported and each element ϕ̂µ(ξ) is localized near the symmetric wedge Wjl = {±ξ, 2j ≤ |ξ| ≤ 2j+1, |θ − θjl| ≤ pi.2−bj/2c}, The number of wedges doubles every other scale doubling and hence the directional selectivity increases for finer scales (see Fig. 5.1). For each µ ∈ M, the curvelet coefficients are given by the inner product cµ = 〈f, ϕµ〉 := ∫ R2 f(x)ϕµ(x)dx (2.8) of a function f ∈ L2(R2) with a curvelet ϕµ ∈ L2(R2). Curvelets are tight frames, so we have the following reconstruction formula f = ∑ µ∈M cµϕµ = CTCf, (2.9) by virtue of the energy conservation ∑ µ∈M |cµ|2 = ‖f‖L2(R2), ∀ f ∈ L2(R2), (2.10) 21 2.2. Approximation of the normal operator with C and CT the curvelet decomposition and composition, respectively. Refer to Candes et al. (2006a), for details on the construction of the fast discrete curvelet transform (FDCT) in dimension two. k1 0-0.25-0.50 0.25 0.50 -0.50 k 2 -0.25 0 0.25 0.500 100 200 300 400 500 0 100 200 300 400 500 Samples S am pl es Figure 2.3: Spatial and frequency representation of curvelets, (a) four dif- ferent curvelets in the spatial domain at three different scales, (b) dyadic partitioning in the frequency domain, where each wedge corresponds to the frequency support of a curvelet in the spatial domain. This figure illustrates the microlocal correspondence between curvelets in the physical and Fourier domain. Curvelets are characterized by rapid decay in the physical space and of compact support in the Fourier space. Notice the correspondence between the orientation of curvelets in the two domains. The 90◦ rotation is a property of the Fourier transform. 2.2.3 Diagonal approximation of PsDO ’s Data, and the scattering operator K can be so defined that the normal operator is a PsDOof order 0, which is real and self-adjoint, and has a homogeneous principal symbol a(x, ξ). We will show that we can approxi- mate Ψf in the curvelet domain by application of a diagonal matrix, if the 22 2.2. Approximation of the normal operator wavenumbers in f are sufficiently high, relative to the smoothness of the symbol of Ψ. So let Ψ = Ψ(x,D) be a pseudodifferential operator of order 0, with homogeneous principal symbol a(x, ξ). Assuming the operator is self-adjoint and real, the principal symbol a is also real and has an even symmetry under the transformation ξ ↔ −ξ. In addition, we make the technical assumption that Ψ is compact, i.e., if f has compact support, then Ψf also has compact support. Curvelets in R2 are denoted by ϕµ and we define |µ| = j. The scale j ranges from some positive number jmin, to infinity (or jmax in an imple- mentation). The central position and wave vector for the curvelet will be denoted by (xµ, ξµ), we let θµ = ξµ/‖ξµ‖ be the angle of the curvelet. Next, we give some consequences of the localization of the curvelet in the space and Fourier domains. The support of a curvelet in the Fourier domain is contained in one of the domains ±ξ2 > |ξ1|, or ±ξ1 > |ξ2|, for some small ε, say we are in the first case, then ξ2 can be used as the radial coordinate, with ξ1/ξ2 the angular coordinate in the Fourier domain. From the localization properties w.r.t. the angular coordinate, it follows that (ξ1/ξ2 − ξµ,1/ξµ,2) is bounded by C12−bj/2c on the support of the curvelet and with C1 some finite positive constant (from here on we assume that any constant Ci is finite and positive), so that ‖(ξ1/ξ2 − ξµ,1/ξµ,2)ϕ̂µ‖L2(R2) ≤ C12−bj/2c. (2.11) Localization in the space domain follows from the smoothness of the defining 23 2.2. Approximation of the normal operator functions in the Fourier domain, in the radial direction we have ‖θµ · (x− xµ)ϕµ‖L2(R2) ≤ C22 −j with C2 another finite constant. For the other directions we have a weaker estimate, as the curvelet is in the Fourier domain narrower, and therefore in the space domain wider in the directions normal to θµ ‖(x− xµ)ϕµ‖L2(R2) ≤ C32 −bj/2c. In fact parallel results hold for curvelets in Rd. We now compare the application of a PsDO to a curvelet with simply multiplying by the value of the symbol at (xµ, ξµ). The result will be proved for x in Rd. Lemma 1 Suppose a is in the symbol class S01,0 (Hormander, 1987), then, with C ′ some constant, the following holds ‖(Ψ(x,D)− a(xν , ξν))ϕν‖L2(Rn) ≤ C ′2−|ν|/2. (2.12) To approximate Ψ, we define the sequence u := (uµ)µ∈M = a(xµ, ξµ) with M the set of all the curvelet indices. Let DΨ be the diagonal matrix with entries given by u. Next we state our result on the approximation of Ψ by CTDΨC. Theorem 1 The following estimate for the error holds ‖(Ψ(x,D)− CTDΨC)ϕµ‖L2(Rn) ≤ C ′′2−|µ|/2, (2.13) 24 2.2. Approximation of the normal operator where C ′′ is a constant depending on Ψ. This main result is proved in Appendix A and it shows that the approxi- mation error for the diagonal approximation decreases for increasingly finer scales, j. The approximation derives from the property that the symbol is slowly varying over the support of a curvelet when the velocity model is sufficiently smooth. 2.2.4 Decomposition of the normal operator By virtue of Theorem 1, the normal operator can approximately be factor- ized into ( Ψϕµ ) (x) ' (CTDΨCϕµ)(x) (2.14) = ( AATϕµ ) (x) with A := CT √ DΨ and AT := √ DΨC. Because the seismic reflectivity can be written as a superposition of curvelets, the ϕµ can be replaced with the model m. The normal equation (cf. Eq. 2.4) can now be rewritten in the following approximate form y(x) = ( Ψm ) (x) + e(x) ' (AATm)(x) + e(x) = Ax0 + e. (2.15) The symbol ' in these expressions is used to indicate that the approxima- tion is only valid for models that are sufficiently close to a reference vector. 25 2.2. Approximation of the normal operator Moreover, the identity only holds to within some constant. This approxi- mation for the forward model forms the point of departure for our seismic amplitude recovery algorithm. To illustrate the above results, a numerical approximation of a PsDO is applied to a number of curvelets at different location, angles and scales. This PsDOcorresponds to the normal operator associated with a seismic experi- ment with a velocity model containing a low-velocity lens (see Fig. 2.4(a)). The results of these experiments are summarized in Fig. 2.2. Fig. 2.2(b), 2.2(d) display the results of applying the PsDO to three different coarse- and fine- scale curvelets (Fig. 2.2(a), 2.2(c)). The Fourier-domain images are plotted in Fig. 2.2(f), 2.2(h). Comparing the imaged curvelets with their originals, shows that curvelets remain invariant as long as they lie in the support of the normal operator. Steep dipping curvelets are outside the support of the operator and this explains the deterioration of their amplitudes for steep dips. As predicted by Theorem 1, the normal operator becomes more diag- onal for increasingly finer scales, an observation reflected in the behavior of the coarse- versus the fine-scale curvelets. Not only is this observation con- sistent with the microlocal correspondence of curvelets reported by Candes and Donoho (2000b) but it is also consistent with the fact that the PsDO ’s correspond to high-frequency approximations of the normal operator. 2.2.5 Estimation of the diagonal The discrete normal operator: Before proceeding with the formulation of the seismic image recovery problem, a method is introduced to estimate the diagonal entries of DΨ. These entries are calculated from applying a 26 2.2. Approximation of the normal operator discrete implementation of the normal operator to some appropriately cho- sen reference vector. This discrete normal operator is given by the com- position of the discretized scattering operator that relates the discretized reflectivity to the discretized data and its adjoint the migration operator. This compound operator corresponds to a matrix-free implementation for Ψ := KTK ∈ RM×M with K ∈ RL×M the scattering matrix and L and M the length of the data and image vectors. The discrete curvelet transform: For the numerical implementation of the curvelet transform, the fast discrete curvelet transform via wrap- ping (Candes et al., 2006a) is used, which corresponds to a matrix-free im- plementation of the tall curvelet decomposition matrix C ∈ RN×M with N = #{M} M with the symbol # denoting ’number of’. This trans- form yields a redundant coefficient vector c = Cr with c ∈ RN . For this choice of curvelet transform, the pseudo-inverse equals the transpose, i.e., CTc = C†c. The transform is a numerical isometry that preserves energy, i.e., ‖r‖ = ‖Cr‖, so we have CTCr = I r in the `2 sense. Since the discrete curvelet representation is overcomplete, with a moderate redundancy (a fac- tor of roughly 8 for d = 2), the converse is not the identity, i.e., CCT r is a projection. Because CCT is a projection, not every curvelet vector is the forward transform of some discrete vector f . Non-uniqueness: The estimation of the diagonal DΨ involves the cal- culation of a diagonal curvelet-domain weighting vector that approximates the action of the normal operator on an appropriately chosen reference vec- 27 2.2. Approximation of the normal operator tor r, i.e., CTDΨCr ' Ψr (cf. Eq. 2.13). This estimation problem can be formulated in terms of a least-squares minimization problem. Define the ’data’ vector as b = Ψr, i.e., the migrated demigrated reference vector. Let DΨ = diag(u), with u ∈ RN , be the diagonal curvelet-domain weighting matrix with coefficients (uµ)µ∈M on its diagonal and calculate the curvelet transform of the reference vector v = Cr. The diagonal estimation problem can now be formulated as finding the vector u that minimizes ‖b − Pu‖2 with P := CT diag(v) and b ∈ RM the known demigrated-migrated refer- ence vector. Since the curvelet transform is redundant (M N), there are many possible solutions to this minimization problem, which gives rise to non-uniqueness. Regularization via smoothness in phase space: One of the conse- quences of the non-uniqueness is the existence of negative entries in the estimated diagonal. These entries are inconsistent with the fact that Ψ is positive definite. Therefore, a regularization is needed that leads to an esti- mate for the diagonal with positive entries only. We argue that this can be accomplished by imposing smoothness in phase space. This smoothness is a property of the symbol that describes the behavior of the normal opera- tor. In the curvelet domain, this smoothness translates to a smoothness in the amplitudes amongst neighboring points on the curvelet grid. For each entry in the diagonal, this smoothness can be imposed by including an ad- ditional penalty term in the least-squares formulation for the estimation of 28 2.2. Approximation of the normal operator the diagonal. This damped least-squares problem can be formulated as ũ = argmin u 1 2 ‖b−Pu‖22 + η2‖Lu‖22, (2.16) where L = [ DT1 D T 2 D T θ ]T is a so-called sharpening operator, penalizing fluctuations amongst neighboring coefficients in u. D1,2 contain first-order differences at each scale in the x1,2 directions, i.e., D1,2 = [ Djmin1,2 · · ·Djmax1,2 ] withDj1,2 the differences at the j th scale in the spatial directions, andDθ the first-order differences in the θ direction, i.e., Dθ = [ Djminθ · · ·Djmaxθ ] with Djθ the difference in the angle direction at the j th scale. These differences are scale dependent because the curvelet grid changes for each scale. The emphasis on the penalty term with respect to the misfit is controlled by η. Estimation with positivity constraints: To ensure a correct balance between the misfit and the penalty functional, Eq. 2.16 is solved for a series of increasing Lagrange multipliers η. For each η, the following system of equations is inverted P ηL u ≈ b 0 (2.17) with ≈ the approximation in least squares sense and 0 a column vector with N zero entries. For a given reference vector, the diagonal estimation procedure consists of the following steps. Apply the normal operator to the reference vector and calculate its curvelet coefficients. Set the Lagrange multiplier to ηmin and solve Eq. 2.16 by inverting the system of equations in Eq. 2.17. Subsequently, increase η by ∆η and repeat this procedure until all 29 2.2. Approximation of the normal operator diagonal elements of ũ are nonnegative. Even though we do not have a proof of convergence for this method, increasing the η leads to positive entries for large enough η. The steps of this estimation procedure are summarized in Table. 2.1. Calculate: b = Ψr and v = Cr. Set: η = ηmin; while ∃ (ũµ)µ∈M < 0 do Solve ũ = argminu 12‖b−Pu‖22 + η2‖Lu‖22 Increase the Lagrange multiplier λ = η +∆η end while Table 2.1: Algorithm for the estimation of the diagonal via regularized least- squares inversion. The Lagrange multiplier, η, is increased up to the point that all entries in the vector for the diagonal are positive. Example: The procedure outlined above is tested on a stylized exam- ple with a velocity model given by a low-velocity lens (Fig. 2.4(a)) and a bandwidth-limited reflectivity (Fig. 2.4(b)), consisting of three events. The low-velocity zone is representative of a gas lens in the overburden. Data is generated from this reflectivity with a zero-order linearized Born modeling and consists of 32 shot records with 500 receivers each. For more details on the migration code, refer to Symes (2006b) and to the numerical ex- periment section at the end of this paper. The data is subsequently imaged (Fig. 2.4(c)). This image and the bandwidth-limited reflectivity (Fig. 2.4(b)) define the ’data’ vector b (cf. Eq. 2.17) and the reference vector r used to estimate ũ according to the algorithm presented in Table 2.1. Fig. 2.5 dis- 30 2.3. Stable seismic image recovery plays the curvelet vectors for increasing η = {0.01, 0.1, 1, 10}. These plots clearly illustrate that ũ becomes positive for η large enough (η ≥ 1 in this case). Fig. 2.4(d) contains the diagonal approximation of the normal opera- tor for the diagonal estimated with η = 1. Comparison between the results of applying the normal operator (Fig. 2.4(c)) and its diagonal approximation (Fig. 2.4(d)), shows a nice correspondence. The relative `2-error between the actual image its approximation is 6.1%. This value corresponds to a moderate size for the constant C ′′ in the bound for the approximation error (Theorem 1). 2.3 Stable seismic image recovery Our formulation for the seismic amplitude recovery problem banks on two fundamental properties of curvelets, namely their ability to detect wave- fronts and their invariance under the normal operator. The combination of these two properties allow for a robust solution of the recovery problem in terms of a nonlinear optimization problem. In this section, the different steps that lead to this formulation are discussed starting with an argumen- tation why curvelets compress seismic images, followed by a quick review of `1-promoted image recovery, its extension to the seismic situation and its solution by a cooling method. 2.3.1 Curvelet frames for seismic images The sedimentary crust consists of sheet-like layers that correspond to func- tions with singularities along piece-wise smooth curves. These singularities 31 2.3. Stable seismic image recovery (a) (b) (c) (d) Figure 2.4: Stylized example of the diagonal estimation for a smooth velocity model and a reflectivity consisting of three reflection events including a fault, (a) the smooth background model with the low-velocity lens. This velocity model is used to calculate the linearized scattering and migration operators, (b) the bandwidth limited reflectivity, (c) the imaged reflectivity from data consisting of 32 shot records with 500 traces each, (d) the approximated normal operator for diag(ũ) estimated with η = 1. The bandwidth limited reflectivity in (b) and the image in (c) served as input for the reference and ’data’ vectors for the diagonal estimation procedure outlined in Table 2.1. Remarks: the main dimming of the normal operator is captured by the diagonal approximation. The relative `2-error between the actual and the approximate normal operator is 6.1%. 32 2.3. Stable seismic image recovery (a) (b) (c) (d) Figure 2.5: Estimates for the diagonal ũ are plotted in (a-d) for increasing η = {0.01, 0.1, 1, 10}. The diagonal is estimated according the procedure outlined in Table 2.1 with the reference and ’data’ vectors, v and b, plotted in Fig. 2.4(b) and 2.4(c). As expected the diagonal becomes more positive for increasing η. 33 2.3. Stable seismic image recovery are associated with rapid variations in the medium fluctuations at which seismic waves reflect. During seismic imaging, these singularities are recov- ered. Examples of singularities imaged from real seismic data can be found in Fig. 2.1(a). This image can be considered as a bandwidth limited version of a function with singularities on piece-wise C2 curves that may include faults and pinch outs (see e.g., Fig. 2.4(b) and 2.8(a) that are examples of synthetic reflectivity models). Fourier (and also wavelet) transforms are known to perform poorly on this type of functions, (Candes and Donoho, 2000a; Candes and Guo, 2002). This poor performance can be explained by the inability of the Fourier transform to localize and the lack of directivity of wavelets. Only when oriented perpendicular to the interfaces, wavelets decay rapidly. Since wavelets lack directionality, they can not resolve these direc- tions, known as the wavefront set. This inability explains why wavelets do not significantly improve the compression rate for seismic images. Curvelet frames, on the other hand, are designed to detect the wavefront set (Candes and Donoho, 2000a, 2004, 2005a,b) and lead, as shown in Fig. 2.6, to better empirical compression rates for real and synthetic seismic images. This im- proved compression rate explains the accurate reconstruction in Fig. 2.1(a) of a real seismic image from only 3% of the curvelet coefficients and is con- sistent with the theoretical nonlinear approximations rates reported in the literature (see e.g. Candes and Donoho, 2000a). These rates correspond to a rate of O(P−2) for curvelets, with P the number of largest entries used in the approximation, while Fourier and wavelets only attain O(P−1/2) and O(P−1), respectively. The empirical rates plotted in Fig. 2.6 seem to confirm these theoretical findings. 34 2.3. Stable seismic image recovery (a) (b) Figure 2.6: Decay rate for the curvelet coefficients compared to the decay for the coefficients of Fourier (dashed line) and discrete wavelet transforms (dot-dashed line), (a) decay rate for the sorted curvelet coefficients (solid line) of the real migrated image included in Fig. 2.1(a), (b) the same as (a) but now for the synthetic reflectivity of the SEG/EAGE AA′ salt model (cf. Fig. 2.8(a)). The decay for the curvelet transform clearly compares favorably to these other two transforms. 2.3.2 Stable recovery Because of the redundancy of the curvelet transform, CCT is a projection and this means that not every curvelet vector is a forward transform of a vec- tor f . Therefore, the vector x0 can’t readily be calculated from f = CTx0, because there exist infinitely many coefficient vectors whose inverse trans- form equals f . Recent work on in the field of compressive sensing has shown that rectangular matrices such as the curvelet matrix, C ∈ RN×M with N M , can stably be inverted through a nonlinear sparsity promoting optimization program. These nonlinear recovery methods require fast decay for the magnitude sorted curvelet coefficients (for details refer to the ex- tensive literature on stable recovery also known as compressed sensing. See e.g. Starck et al., 2004; Elad et al., 2005; Candes et al., 2006b). Following these results, the vector x0 can be approximately recovered 35 2.3. Stable seismic image recovery from noise-corrupted data y = Ax0 + n (2.18) with A := CT . As long as the signal is sufficiently compressible x0 can successfully be recovered to within the noise level. This recovery involves the solution of a constrained convex optimization problem P1 : x̃ = argminx ‖x‖1 := ∑N j=1 |xj | subject to ‖y −Ax‖2 ≤ m̃ = Ax. (2.19) For certain matricesA, the solution of this unconstrained optimization prob- lem lies to within the noise level (see e.g. Candes et al., 2006b; Elad et al., 2005). The optimization problem P1 is known as the constrained variation of the LASSO (Tibshirani, 1997) and the basis-pursuit denoising (BPDN) (Chen et al., 2001) algorithms. As part of the optimization, the sparsity vector is fitted within the tol- erance . This tolerance depends on the noise level given by the standard deviation of the noise vector. Since n1, n2, · · ·nM ∈ N(0, σ2), the probabil- ity of ‖n‖22 exceeding its mean by plus or minus two standard deviations is small. The ‖n‖22 is distributed according the χ2-distribution with mean M ·σ2 and standard deviation √2M ·σ2. By choosing 2 = σ2(M + ν√2M) with ν = 2, we remain within the mean plus or minus two standard devi- ations. Following (Elad et al., 2005), the above constrained optimization problem (P1), is replaced by a series of simpler unconstrained optimization 36 2.3. Stable seismic image recovery problems Pλ : argminx ‖y −Ax‖22 + λ‖x‖1 m̃ = Ax̃. (2.20) These optimization problems depend on the Lagrange multiplier λ. A cool- ing method is used, where Pλ is solved for a Lagrange multiplier λ that is slowly decreased from a large starting value λ1 < ‖ATy‖∞. The optimal x̃ is found for the largest λ for which ‖y−Ax̃‖2 ≤ . During the optimization, the underdetermined frame matrix A is inverted by imposing the sparsity promoting `1-norm. This norm regularizes the inverse problem of finding the unknown coefficient vector (see also Daubechies et al., 2005). We refer to (Donoho et al., 2006; Tropp, 2006) for the recovery conditions for Eq.’s 2.19 and 2.20. 2.3.3 Solution by iterative thresholding Following (Daubechies et al., 2005; Elad et al., 2005) and ideas dating back to (Figueiredo and Nowak, 2003), Eq. 2.27 is solved by an iterative thresholding technique that derives from the Landweber descent method. The method consist of an outer loop during which the Lagrange multiplier is lowered and an inner loop during which Pλ is solved. The inner loop is initialized by the solution obtained from the previous outer loop starting with zero vector for the first iteration. After m iterations of the outer cooling loop, the estimated coefficient vector is computed for fixed λ by the following inner loop xm+1 = Tλ ( xm +AT (y −Axm)) (2.21) 37 2.3. Stable seismic image recovery with λ = λm and Tλ(x) := sgn(x) ·max(0, |x| − |λ|). (2.22) As shown by (Daubechies et al., 2005), this iteration for fixed λ converges to the solution of Eq. 2.27 for m large enough and ‖A‖ < 1. The cost for each iteration is a curvelet synthesis and subsequent analysis. 2.3.4 Stable seismic recovery The main purpose is to estimate the relative amplitudes of seismic images from data that are possibly contaminated with noise. This data is repre- sented by the discretized linearized forward model d = Km+ n (2.23) withK the discrete linearized Born modeling operator. Applying the adjoint of this operator to the data vector creates a noisy image (cf. Eq. 2.4) y = Ψm+ e. (2.24) With the curvelet decomposition of the normal operator (cf. Eq. 3.9) in discrete form, i.e., Ψm ' AATm, (2.25) 38 2.3. Stable seismic image recovery this expression can be rewritten into the following approximate form y ' Ax0 + e (2.26) with A := CTΓ and Γ := √ DΨ. To avoid instabilities related to small entries in the estimated diagonal u, we set DΨ = diag (ũ + δ)/δ with δ some small parameter. Comparing this image representation with the one in Eq. 2.18 shows that these expressions are equivalent, aside from the noise term. While the noise in Eq. 2.18 is white and Gaussian, the noise term in Eq. 2.26 is colored by the migration operator. The expectation for the covariance of this colored noise term equals E{eeT } = σ2Ψ ' σ2AAT . Because the nonlinear recovery of the vector x0 involves the inversion of the matrix A in which the covariance is factored, the noise term is approxi- mately whitened. To understand this whitening, consider the pseudo-inverse A† = Γ−1C which, when applied to e, approximately whitens the noise in the curvelet domain. This whitening is a well-known property of WVD techniques. The nonlinear recovery of seismic images now involves solving P′1 : x̃ = argminx ‖x‖1 := ∑N j=1 |xj | subject to ‖y −Ax‖2 ≤ m̃ = ( AT )† x̃. (2.27) The main assumption underlying this formulation is that the curvelet vector of the model remains (approximately) sparse after application of the normal operator. This assumption is valid when the curvelets remain sufficiently 39 2.4. Sparsity- and continuity-enhanced seismic image recovery invariant under this operator. In that case, the nonlinear program (P′1) approximately inverts the normal operator in two stages, namely during the `1-norm regularized inversion of the synthesis matrix, yielding a sparse estimate for the curvelet coefficient vector, and during the calculation of the model vector via the pseudo-inverse of AT . During each stage, the ’square-root’ of the normal operator is approximately inverted. 2.4 Sparsity- and continuity-enhanced seismic image recovery The above formulation provides a stable recovery for the amplitudes of seis- mic images, the relative strengths of the fluctuations in m, without the necessity of multiple evaluations of the normal operator. This recovery is only accurate when the proposed diagonal approximation (cf. Eq. 3.3) pro- vides an adequate approximation to the normal operator. The accuracy of this approximation depends on the available scales in the seismic image (cf. Theorem. 1), on the complexity of the background velocity model, the acquisition geometry and the closeness of the reference r to the actual un- known modelm. For the remainder of this paper, we assume the acquisition to be ideal and the reference vector to be close to the actual model. Anisotropic diffusion: Aside from the aforementioned factors that enter into the seismic recovery problem, there is the issue of how to limit spurious curvelet artifacts. These artifacts are either related to the so-called “pseudo Gibb’s” phenomenon (or better side-band effects (see Candes and Guo, 40 2.4. Sparsity- and continuity-enhanced seismic image recovery 2002)), inherent to curvelet or other harmonic expansions. To reduce these spurious artifacts, the sparsity enhancing penalty functional in Eq. 2.27 is complemented with an additional continuity-enhancing penalty functional. This functional enhances the continuity along the imaged reflectors, which tend to be smooth in the tangential directions. By applying an anisotropic smoothing technique, the ’wavefront’ set of the imaged reflectivity is pre- served. This technique differs from commonly used edge-preserving penalty functionals such as total variation (TV) (see e.g. Claerbout and Muir, 1973; Schertzer, 2003) that tend to remove the oscillatory behavior of the reflectors in the normal direction. The anisotropic-diffusion penalty term (see e.g. Fehmers and Hocker, 2003) is given by Jc(m) = ‖Λ1/2∇m‖22, (2.28) with ∇ the discretized gradient matrix defined as ∇ = [DT1 DT2 ]T . The block-diagonal matrix Λ is location dependent (see Fig. 2.10, which plots the gradients) and rotates the gradient towards the tangents of the reflecting surfaces. This rotation matrix is given by Λ[r] = 1 ‖∇r‖22 + 2υ +D2r −D1r (+D2r −D1r)+ υI , (2.29) with Di the discretized derivative in the ith coordinate direction and υ a parameter that controls the fluctuations for regions where the gradient is small. Following Black et al. (1998), this control parameter is set propor- tional to the median of |∇r| with | | the length of each gradient vector (white 41 2.4. Sparsity- and continuity-enhanced seismic image recovery arrows in Fig. 2.10). Similar to the diagonal approximation, a reference vec- tor derived from the migrated image (cf. Eq. 2.24) is used to calculate the tangential directions of the reflecting surfaces. By combining the two different penalty terms that promote sparsity and continuity, we finally arrive at our formulation for the seismic-amplitude recovery problem P : argminx J(x) subject to ‖y −Ax‖2 ≤ m̃ = ( AT )† x, (2.30) in which the composite penalty term J(x) is given by J(x) = αJs(x) + βJc(x), (2.31) with α, β ≥ 0 and α+β = 1. The Js(x) = ‖x‖1 is the `1-norm. The second term in the penalty term is given by Jc(x) = ‖Λ1/2∇ ( AT )† x‖22. Because the optimization is carried out over x and not over the model vectorm, this expression includes a pseudo-inverse that is calculated with a few iterations of the LSQR algorithm (Paige and Saunders, 1982). The cooling method: The above nonlinear optimization problem (P) is solved with a cooling method (see e.g. Starck et al., 2004). This method consists of a series of thresholded Landweber iterations that solve a series of unconstrained subproblems for decreasing λ. Since this method only requires knowledge of the Jacobians at each iteration, it is relatively straightforward to include the Jacobian of the additional continuity-enhancing penalty func- 42 2.4. Sparsity- and continuity-enhanced seismic image recovery tional Jc(x). For data given by Eq. 2.18, the iterations of the cooling method for a particular cooling parameter λ consist of the following three main steps: Step 1: update of the Jacobian of 12‖y −Ax‖22: x← x+AT (y −Ax) ; (2.32) Step 2: project onto the `1 ball S = {‖x‖1 ≤ ‖x0‖1} by soft thresholding x← Tλ(x); (2.33) Step 3: project onto the anisotropic diffusion ball C = {x : J(x) ≤ J(x0)} by x← x− β∇xJc(x) (2.34) with ∇xJc(x) = 2A†∇ · ( Λ∇ (( AT )† x )) . (2.35) Remember that steps 1&2 with ‖A‖ ≤ 1 converge to the solution of Eq. 2.27 for a fixed λ. During step 3, the coefficients are updated according to the gradient of the anisotropic diffusion norm designed to reduce spurious arti- facts. The different steps of our algorithm are summarized in Table 2.2. To ensure sparsity, the algorithm is started by setting x0 = 0. 43 2.5. Practical considerations Initialize: m = 0; x0 = 0; y = KTd; Choose: M and L ‖ATy‖∞ > λ1 > λ2 > · · · while ‖y −Ax̃‖2 > do m = m+ 1; xm = xm−1; for l = 1 to L do xm = Tλm ( xm +AT (y − xm) ){Iterative thresholding} end for Anisotropic descent update; xm = xm − β∇xm Jc(xm); end while x̃ = xm; m̃ = ( AT )† x̃. Table 2.2: Sparsity-and continuity-enhancing recovery of seismic amplitudes. 2.5 Practical considerations The presented method depends on the availability of a reference vector with an image that is sufficiently close to the unknown model. As in most Krylov- subspace solvers, we derive the reference vector from the migrated image (matched filter) as an initial guess. Since waves are subject to spherical spreading, a depth-dependent correction is applied to the migrated image (cf. Eq. 2.24). This corrected image serves as input for the diagonal estima- tion procedure and the calculation of the anisotropic difussion norm. The image recovery procedure consists of the following steps 1. Obtain the reference vector, r, by imaging the data with Eq. 2.24, 44 2.6. Numerical experiments followed by applying the depth correction. 2. Use the reference vector, r, to estimate the diagonal matrix, diag(ũ), with the procedure outlined in Table 2.1; 3. Define the synthesis and analysis matrices A and AT ; 4. Calculate the rotation matrix, Λ, (cf. Eq. 2.29) for the continuity promoting penalty functional; 5. Solve the optimization problem P (Eq. 2.30) with the algorithm out- lined in Table 2.2. This yields the estimate for the amplitude-corrected image, m̃; 6. Optionally goto 1 with r := m̃; 2.6 Numerical experiments The recovery method is applied to a synthetic imaging example for data of the SEG/EAGE AA′ salt model (O’Brien and Gray, 1997; Aminzadeh et al., 1997). The original velocity model and its smoothed version that defines the background velocity model, used by the migration operators, are included in Fig. 2.7. To avoid multiple reflections, the background velocity model is averaged over a window of size 720 × 720m. The reflectivity (velocity perturbation) is defined as the band-pass filtered difference between the smoothed velocity model and a slightly less smoothed velocity model. The density is assumed to be constant. Fig. 2.8(a) shows the reflectivity model that is 3.6 km by 15.3 km with a grid spacing of 24m in the vertical and 45 2.6. Numerical experiments horizontal coordinate directions. This reflectivity was used to generate data with the linearized Born modeling operator. The migrated image: A two-way reverse-time migration code with check- pointing was used to conduct the experiments. This two-dimensional (d = 2) time-domain code, developed by Symes (2006b), has absorbing boundary conditions and is based on the adjoint-state method. The migration oper- ator is derived from the gradient of a descent algorithm for least-squares inversion (see Symes, 2006b, and the references therein). The data is gen- erated with the adjoint of the migration operator, which corresponds to the linearized Born approximation. This Born modeling operator generates data without nonlinear events such as multiple reflections. The migration operator and its adjoint are made zero-order by applying a half-order frac- tional integration on the source wavelet. We choose a source wavelet with a relatively broad temporal frequency band [5 − 60]Hz. The data consists of 324 shots with 176 receivers each with a shot at every 48m, yielding a maximum offset of 4.224 km. Each geophone records 626 time samples with a sample interval of 8ms, which amounts to 5 s of data. The 8000 time-step linearized Born modeling takes about 64min with 68 CPU’s on a MPI cluster, while the migration takes about 294min. These computation times explain why calculating the pseudo-inverse of the normal operator becomes computationally prohibitive in real applications where the images are three dimensional (d = 3). Fig. 2.8 shows the original reflectivity and the migrated image obtained with Eq. 2.24. The image suffers from amplitude deterioration due to spher- 46 2.6. Numerical experiments (a) (b) Figure 2.7: The SEG/EAGE AA′ salt model (Aminzadeh et al., 1997), (a) The original velocity model, (b) the smoothed velocity model with a window size of 720 × 720m. Remark: Imaging this model is a challenge because of the presence of the high-velocity (bright) salt. 47 2.6. Numerical experiments ical spreading and poor illumination. Despite the overall amplitude deteri- oration for steep reflectors and events at increasing depths, the migration resolves most of the singularities in the image (cf. Fig 2.8(a) and 2.8(b)). (a) (b) Figure 2.8: Imaging of the SEG/EAGE AA′ salt model (Aminzadeh et al., 1997), (a) reflectivity defined in terms of the band-pass filtered difference between the smoothed velocity model and a slightly less smoothed velocity model, (b) the imaged reflectivity according to Eq. 2.24. The main reflec- tion events are present but suffer from deteriorated amplitudes, especially under the high-velocity salt and for steep reflectors and faults. 48 2.6. Numerical experiments Diagonal estimation: After applying corrections for the spherical spread- ing, the migrated image is considered as reference vector. This reference vector is used to compute the diagonal approximation for the normal op- erator and for the calculation of the gradients. The image deterioration between the reference vector (Fig. 2.9(a)) and the demigrated-migrated ref- erence vector (Fig. 2.9(b)) is similar to the amplitude deterioration between the ’unknown’ true reflectivity and migrated image (cf. Fig. 2.8), there- for the images in Fig’s. 2.9(a) and 2.9(b) should be adequate as input for the diagonal estimation procedure outlined in Table 2.1. To avoid insta- bilities related to small entries in the estimated diagonal diag ũ, we set Γ = diag (ũ + δ)/δ with δ = 0.2. The result of applying the estimated diagonal to the bandwidth-limited reflectivity of Fig. 2.8(a) is included in Fig. 2.9(c) and shows that most of the imprint of the normal operator is captured by the diagonal approximation. Seismic amplitude recovery for the SEG AA′ model: With the di- agonal approximation in place, the amplitude corrections are applied with the procedure presented in Table 2.2. The anisotropic diffusion penalty term is calculated from the gradients plotted in Fig. 2.10. The results of the amplitude recovery are summarized in Fig. 2.11. Fig. 2.11(a) depicts the result obtained by only applying a `1-norm penalty (β = 0). The image is calculated by lowering the Lagrange multiplier λ in 20 steps from a value that corresponds to a threshold that removes all but 5% of the coefficients to a Lagrange multiplier that only removes 1% of the curvelet coefficients. For each intermediate λ, the inner loop is repeated 5 times (L=5). The 49 2.6. Numerical experiments (a) (b) (c) Figure 2.9: Images that are input to the diagonal approximation scheme outlined in Table 2.1, (a) the reference vector, r, derived from Fig. 2.8(b), after applying corrections for the spherical-divergence and receiver-array ta- per, (b) the demigrated-migrated reference vector, b = Ψr, that serves as ’data’ for the diagonal estimation, (c) diagonal approximation on the original bandwidth-limited reflectivity plotted in (a). This diagonal approx- imation by AATm captures the normal operator, Ψm, quite well. 50 2.6. Numerical experiments Gradient of the reference vector lateral (m) de pt h (m ) 2000 4000 6000 8000 10000 12000 14000 500 1000 1500 2000 2500 3000 3500 Figure 2.10: Gradient vectors for the reference vector r plotted in Fig. 2.9(a). The gradients (white ↑’s) are used to calculate the tangential di- rections along which the additional anisotropic smoothing is applied. 51 2.6. Numerical experiments result of this norm-one optimization are plotted in Fig. 2.11(a). The im- age shows a clear improvement over the image plotted in Fig. 2.9(a). Not only is the spherical divergence corrected in a stable manner, but the ampli- tude of the steep reflections events are recovered as well. There are, however, some small remaining artifacts related to side-band effects (Candes and Guo, 2002). By including the continuity-enhancing penalty functional most of the aforementioned artifacts can be removed as can be seen from Fig. 2.11(b). In this example, the sparsity- and continuity-enhancing norms are weighted equally, i.e. α = β = 1/2 in Eq. 2.31. To further illustrate the migration amplitude recovery, detailed plots for different traces of the recovered the amplitudes are included in Fig. 2.12. Fig. 2.12(a) compares vertical traces of the original, migrated and amplitude recovered images. The traces are normalized with respect to the first reflector and the plots show that the amplitudes are nicely recovered. Similarly, the amplitudes along the hori- zontal reflector at z = 3432m, are nicely recovered. The improvement in the recovered amplitudes is also apparent from Fig. 2.12(c) that contains the average vertical-wavenumber amplitude spectra of the original, the imaged and the recovered reflectivity. Seismic amplitude recovery from noisy data: So far, the examples were noise free. In practice, seismic data volumes contain several sources of clutter, ranging from coherent nonlinear events, such as multiple reflections, to incoherent measurement errors. In this paper, only the recovery from incoherent contamination is considered. Sources of coherent clutter can be removed by using curvelet-based signal separation techniques reported 52 2.6. Numerical experiments (a) (b) Figure 2.11: Images after nonlinear recovery (P). (a) result with the `1- norm recovery only (β = 0); (b) recovery with the combined `1 and con- tinuity recovery for α = β = 1/2. The amplitudes are recovered. The anisotropic diffusion successfully removes the artifacts. 53 2.6. Numerical experiments (a) (b) (c) Figure 2.12: Recovery of the amplitudes after normalization with respect to the first event, (a) seismic traces of the original reflectivity (dot-dashed line), the migrated reflectivity (dashed line) and the recovered reflectivity (solid line), (b) amplitudes along the bottom most horizontal reflector, (c) average amplitude spectra of the depth-dependent reflectivity. The original and recovered spectra are normalized to match. Remarks: The nonlinear recovery corrects for the amplitudes and restores the spectrum. 54 2.6. Numerical experiments elsewhere (Herrmann et al., 2007). An image computed from data with a signal-to-noise ratio (SNR) of 3 dB is included in Fig. 4.3(a). As shown in Eq. 2.24, the image now contains an additional contribution due to the noise. This noise contribution is clearly visible throughout the image as a non-stationary clutter and leads to a deterioration of the image quality. The clutter in the image space, however, is significantly smaller than in the data space, as can be observed from the noisy shot record plotted in Fig. 2.14(b). The difference in noise levels between the data and image spaces can be explained by the 154-fold redundancy of the data space compared to the image space (L = 154×M). The result for the amplitude recovery of the noisy data are included in Fig. 4.3(b) and this figure shows that the migration amplitudes can stably be recovered, leading to a significant improvement in the SNR for the image (9.2 dB for the image in Fig. 4.3(b) as opposed to (1.5 dB for the noisy image which includes the error due to the imprint of the normal operator). This improved image was obtained using the algorithm of table 2.2, with the same settings as in the noise-free example except for the lowest value of the Lagrange multiplier, which is now set to a larger value depending calculated from the noise level. From Fig. 4.3(b), one can see that the noise contamination has largely been removed. Moreover, the amplitudes have been restored in particular for the steep events at depth. Data generated from the estimated image, d̃ = Km̃ shows a significant removal of the noise (cf. Fig. 2.14(b)-2.14(c)), with reflection events that match the noise- free data plotted in Fig. 2.14(a). This visual improvements leads to an improvement of SNR for the data of 19.2 dB. 55 2.7. Conclusions 2.7 Conclusions 2.7.1 Initial findings The method presented in this paper combines the compression of images by curvelets with the invariance of this transform under the normal operator. This combination allows for a formulation of a stable recovery method for seismic amplitudes. During the recovery, the normal operator is approxi- mately inverted. Compared to other approaches for migration scaling, the presented method (i) includes a theoretical bound on the L2 − error for the diagonal approximation in the curvelet domain; (ii) prescribes a proce- dure for the estimation of the diagonal from numerical implementations of the imaging operators; (iii) formulates the amplitude recovery problem as a nonlinear optimization problem, where the inversion of the diagonalized normal operator is regularized by imposing sparsity in the curvelet domain and continuity along the imaged reflectors. As long as the background velocity model is sufficiently smooth and there is a reference vector available close enough to the actual reflectivity, the normal operator can be approximated by a diagonal weighting in the curvelet domain. The theoretical approximation error is scale-dependent and decreases for finer scales, an observation consistent with the microlocal properties of the normal operator and curvelets in the high-frequency limit. Estimation of the diagonal exploits smoothness of the symbol by penalizing neighboring entries in the diagonal that fluctuate. These fluctuations are inconsistent with the smoothness of the symbol. For a sufficiently large La- grange multiplier, the regularized least-squares estimation for the diagonal 56 2.7. Conclusions (a) (b) Figure 2.13: Image amplitude recovery for noisy data (SNR 3dB), (a) noise image according to Eq. 2.24, (b) image after nonlinear recovery from noisy data (P). The clearly visible non-stationary noise in (a) is removed dur- ing the recovery while the amplitudes are also restored. Steeply dipping reflectors denoted by the arrows are also well recovered. 57 2.7. Conclusions (a) (b) (c) Figure 2.14: ’Denoising’ of a shot record, (a) the noise-free data received by the receiver array, (b) noisy data with a SNR of 3 dB, (c) forward modeled data after amplitude recovery. Observe the significant improvement in the data quality, reflected in an increase for the SNR of 19.2 dB. 58 2.7. Conclusions leads to positive entries. Numerical experiments with the diagonal approx- imation showed a moderate-sized constant for the bound of the theoretical error. The diagonal approximation is used to formulate the seismic amplitude recovery in terms of a constrained optimization problem. The amplitude- corrected image is obtained by solving this sparsity- and continuity-promoting optimization problem. The invariance of curvelets under the normal opera- tor preserves the sparsity. The cost of computing the diagonal approxima- tion is one demigration-migration per reference vector, which is much less compared to the cost of Krylov-based least-squares inversion. The recov- ery results show an overall improvement of the image quality. The joined sparsity- and continuity-enhanced image has diminished artifacts, improved resolution and recovered amplitudes. 2.7.2 Extension to prestack imaging The discussion has been on ’post-stack’ images obtained after applying the zero-offset imaging condition. When the velocity model is correct, images are smooth along the redundant coordinate in pre-stack angle gathers. This smoothness property has been used in velocity analyses methods such as differential semblance (Symes and Carazzone, 1991) and is suitable for the framework laid down in this paper. Besides the `1- and continuity norms, the amplitude recovery can also exploit the smoothness along the redundant angle coordinate. The fact that smoothness translates to sparsity is espe- cially exciting because it allows to further exploit the results from the field of compressed sensing. 59 2.7. Conclusions 2.7.3 Recent related work During the revision, we were informed of recent work by William Symes on optimal scaling of reverse-time migration Symes (2006a). This work and recent work by Guitton (2004) attempt to invert the normal operator by estimating a diagonal approximation from a reference vector, typically given by the migrated image, and the demigrated-migrated reference vec- tor. Smoothness of the diagonal approximation is also imposed by both authors by parameterizing the diagonal as a smooth function in the spatial domain. Our approach goes several steps further by imposing smoothness in phase space and by making use of the curvelet transform that allows for the inversion of the diagonal using the methods of stable image recovery. 2.7.4 Choice of the reference vector Having access to a proper reference vector is a prerequisite for the success of the method presented in this paper. Selection of the appropriate reference is somewhat reminiscent of finding a good initial guess for Krylov subspace methods. In this paper, we took a data-driven approach by using the mi- grated image (matched-filter result) as a first guess. Our formulation allows for multiple reference vectors and we hope to report on appropriate selection strategies for reference vectors in the future. This investigation will include an analysis of the sensitivity of our method to the accuracy of the velocity model. 60 2.7. Conclusions Acknowledgments The authors would like to thank the reviewers for their constructive com- ments and suggestions. We also wish to thank the authors of the Fast Dis- crete Curvelet Transform for making their code available at www.curvelet.org and Dr. William Symes for making his reverse-time migration code avail- able. M. O’Brien, S. Gray and J. Dellinger from BP Amoco are thanked for making the SEG AA’ dataset available. The examples presented in this paper were prepared with CWP’s Seismic Un*x and with Madagascar (rsf.sourceforge.net). This work was in part financially supported by the Natural Sciences and Engineering Research Council of Canada Discovery Grant (22R81254) and the Collaborative Research and Development Grant DNOISE (334810-05) of F.J.H. This research was carried out as part of the SINBAD project with support, secured through ITF (the Industry Technol- ogy Facilitator), from the following organizations: BG Group, BP, Chevron, ExxonMobil and Shell. The first two authors would also like to thank the Institute of Pure and Applied Mathematics at UCLA, supported by the NSF under grant DMS-9810282, for their hospitality. The research of the third author was supported by the Netherlands Organization for Scientific Research, through a VIDI grant no. 639.032.509. 61 Bibliography Aminzadeh, F., J. Brac, and T. Kunz, 1997, 3-D salt and overthrust model. SEG/EAGE 3-D Modeling Series, No. 1: Society of Exploration Geo- physicists, Tulsa. Beylkin, G., 1984, The inverse problem and applications of the generalized Radon transform: Comm. Pure Appl. Math., 37, 579–599. Black, M., G. Sapiro, and D. Marimont, 1998, Robust anisotropic diffusion: IEEE Trans. Signal Process., 7, 421–432. Bleistein, N., J. Cohen, and J. Stockwell, 2001, Mathematics of multidimen- sional seismic imaging, migration and inversion: Springer. Brandsberg-Dahl, S. and M. V. de Hoop, 2003, Focusing in dip and AVA compensation on scattering-angle/azimuth common image gathers: Geo- physics, 68, 232–254. Candes, E. and D. Donoho, 2005a, Continuous curvelet transform i: Reso- lution of the wavefront set: Appl. Comput. Harmon. Anal., 19, 162–197. Candes, E. J., L. Demanet, D. L. Donoho, and L. Ying, 2006a, Fast discrete curvelet transforms: SIAM Multiscale Model. Simul., 5, 861–899. Candes, E. J. and D. Donoho, 2004, New tight frames of curvelets and optimal representations of objects with piecewise C2 singularities: Comm. Pure Appl. Math., 57, 219–266. 62 Chapter 2. Bibliography ——–, 2005b, Continuous curvelet transform ii: Discretization and frames: Appl. Comput. Harmon. Anal., 19, 198–222. Candes, E. J. and D. L. Donoho, 2000a, Curvelets – a surprisingly effec- tive nonadaptive representation for objects with edges: Presented at the Curves and Surfaces, Vanderbilt University Press. ——–, 2000b, Recovering edges in ill-posed problems: Optimality of curvelet frames: Ann. Statist., 30, 784. Candes, E. J. and F. Guo, 2002, New multiscale transforms, minimum total variation synthesis: Applications to edge-preserving image reconstruction: Signal Processing, 1519–1543. Candes, E. J., J. Romberg, and T. Tao, 2006b, Stable signal recovery from incomplete and inaccurate measurements: Comm. Pure Appl. Math., 59, 1207–1223. Chavent, G. and R. Plessix, 1999, An optimal true-amplitude least-squares prestack depth-migration operator: Geophysics, 64, 508–515. Chen, S. S., D. L. Donoho, and M. A. Saunders, 2001, Atomic decomposition by basis pursuit: SIAM Journal on Scientific Computing, 43, 129–159. Claerbout, J. and F. Muir, 1973, Robust modeling with erratic data: Geo- physics, 38, 826–844. Claerbout, J. and D. Nichols, 1994, Spectral preconditioning: Technical report, Stanford Exploration Project. Daubechies, I., M. Defrise, and C. de Mol, 2005, An iterative thresholding algorithm for linear inverse problems with a sparsity constraints: Comm. Pure Appl. Math., 57, 1413–1457. de Hoop, M. V. and N. Bleistein, 1997, Generalized Radon transform inver- 63 Chapter 2. Bibliography sions for reflectivity in anisotropic elastic media: Inverse Problems, 13, 669–690. de Hoop, M. V. and S. Brandsberg-Dahl, 2000, Maslov asymptotic extension of generalized Radon transform inversion in anisotropic elastic media: a least-squares approach.: Inverse problems, 16, 519–562. Donoho, D. L., 1995, Nonlinear solution of linear inverse problems by wavelet-vaguelette decomposition: App. and Comp. Harmonic Analysis, 2, 101–126. Donoho, D. L., M. Elad, and V. Temlyakov, 2006, Stable recovery of sparse overcomplete representations in the presence of noise: IEEE Trans. In- form. Theory, 52, 6–18. Elad, M., J. L. Starck, P. Querre, and D. L. Donoho, 2005, Simultane- ous cartoon and texture image inpainting using morphological component analysis (mca): Appl. Comput. Harmon. Anal., 19, 340–358. Fehmers, G. C. and C. F. Hocker, 2003, Fast structural interpretation with structure-oriented filtering: Geophysics, 68, 1286–1293. Figueiredo, M. and R. Nowak, 2003, An em algorithm for wavelet-based image restoration: IEEE Trans. Image Processing. Guitton, A., 2004, Amplitude and kinematic corrections of migrated images for nonunitary imaging operators: Geophysics, 69, 1017–1024. Hennenfent, G. and F. J. Herrmann, 2006, Seismic denoising with non- uniformly sampled curvelets: Comp. in Sci. and Eng., 8, 16–25. Herrmann, F. J., 2003, Multifractional splines: application to seismic imag- ing: Proceedings of SPIE Technical Conference on Wavelets: Applications in Signal and Image Processing X, 240–258. 64 Chapter 2. Bibliography Herrmann, F. J., U. Boeniger, and D. J. Verschuur, 2007, Nonlinear primary- multiple separation with directional curvelet frames: Geoph. J. Int., 170, 781–799. Hormander, L., 1987, The analysis of linear partial differential operators iii: Pseudo-differential operators. Hu, J., G. T. Schuster, and P. Valasek, 2001, Poststack migration deconvo- lution: Geophysics, 66, 939–952. Kuhl, H. and M. Sacchi, 2003, Least-squares wave-equation migration for AVP/AVA inversion: Geophysics, 68, 262. Lee, N. Y. and B. J. Lucier, 2001, Wavelet methods for inverting the Radon transform with noisy data: IEEE Trans. Image Processing, 10, 79–94. Nemeth., T., C. Wu, and G. T. Schuster, 1999, Least-squares migration of incomplete reflection data: Geophysics, 64, 208–221. O’Brien, M. and S. Gray, 1997, Can we image beneath salt?: Leading Edge, 15, 17–22. Paige, C. C. and M. A. Saunders, 1982, Lsqr: An algorithm for sparse linear equations and sparse least squares: ACM TOMS, 8, 43–71. Plessix, R. and W. Mulder, 2004, Frequency-domain finite difference amplitude-preserving migration: Geoph. J. Int., 157, 975–987. Rickett, J. E., 2003, Illumination-based normalization for wave-equation depth migration: Geophysics, 68, no. 4. Schertzer, O., 2003, Scale-space methods and regularization for denosing and inverse problems: Advances in Imaging and Electron Physics, 128, 445–530. Starck, J. L., M. Elad, and D. L. Donoho, 2004, Redundant multiscale 65 Chapter 2. Bibliography transforms and their application to morphological component separation: Advances in Imaging and Electron Physics, 132. Stolk, C. C., 2000, Microlocal analysis of a seismic linearized inverse prob- lem: Wave Motion, 32, 267–290. Stolk, C. C. and W. W. Symes, 2003, Smooth objective functionals for seismic velocity inversion: Inverse Problems, 19, 73–89. Symes, W. W., 2006a, Optimal scaling for reverse time migration: Technical Report TR 06-19, Department of Computational and Applied Mathemat- ics, Rice University, Houston, Texas, USA. ——–, 2006b, Reverse time migration with optimal checkpointing: Technical Report 06-18, Department of Computational and Applied Mathematics, Rice University, Houston, Texas, USA. Symes, W. W. and J. Carazzone, 1991, Velocity inversion by differential semblance optimization: Geophysics, 56, 2061–2073. ten Kroode, A., D.-J. Smit, and A. Verdel, 1998, A microlocal analysis of migration: Wave Motion, 28. Tibshirani, R., 1997, Regression shrinkage and selection via the LASSO: J. Royal. Statist. Soc B., 58, 267–288. Tropp, J. A., 2006, Just relax: convex programming methods for identifying sparse signals in noise: IEEE Trans. Inform. Theory, 52, 1030–1051. Tsaig, Y. and D. L. Donoho, 2006, Extensions of compressed sensing: Signal processing, 86, 549–571. Ying, L., L. Demanet, and E. J. Candes, 2005, 3D discrete curvelet trans- form: Presented at the Wavelets XI, SPIE. 66 Chapter 3 Curvelet-based migration preconditioning and scaling 3.1 Introduction Over the years, extensive research has been done to reduce the computa- tional costs of (least-squares) seismic imaging. Improvements in this area are particularly important during iterative least-squares migration, where the linear Born-scattering operator is inverted with iterative Lanczos meth- ods, such as LSQR (Paige and Saunders, 1982; De Roeck, 2002). Examples of these methods can be found in the literature (see e.g. Nemeth et al., 1999; Chavent and Plessix, 1999; Hu et al., 2001; Kuhl and Sacchi, 2003; Yu et al., 2006). The most successful methods to reduce the cost of migration are the so- called scaling methods where the action of the compound linearized modeling- migration operator—known as the Hessian or normal operator—is replaced by a diagonal scaling in some domain, see e.g. contributions by Claerbout A version of this chapter has been published. Herrmann, F.J., Brown, C.R., Erlangga, Y.A. and Moghaddam, P.P. (2009) Curvelet-based migration preconditioning and scaling. Geophysics, 74, pp. A41, 2009 67 3.1. Introduction and Nichols (1994); Rickett (2003); Guitton (2004); Plessix and Mulder (2004), and more recently by Symes (2008) and Herrmann et al. (2008a). These methods vary in degree of sophistication with regard to the estimation of the diagonal, e.g. through migrated-image to remigrated-image match- ing. They also differ in the way the scaling is applied—i.e., by ’division’ in the physical domain or via sparsity promotion in the curvelet domain, as reported recently by Herrmann et al. (2008a). During all these methods, im- aged amplitudes are restored by applying a scaling as a post-processing step after migration. Even though our approach and the one of Symes (2008) are similar, the curvelet-domain scaling has the advantage that it is able to handle conflicting dips. In this paper, we take this line of research a step further by using the above scaling argument to apply the proper preconditioning to the system of equations involved in linearized Born scattering. By “proper” we mean a preconditioner with a computational overhead that justifies its improvement by the increase in convergence rate. Because the system is solved iteratively, the preconditioner does not need to be very accurate, avoiding unnecessary extra computational overhead. Note that we use the term precondition- ing somewhat loosely compared to its formal definition in numerical linear algebra where the solution is not changed with preconditioning (Calvetti, 2007). Therefore, we also use this term to denote changes in the forward model that favor least-squares inversion. To illustrate the improvements in the images and in the convergence of least-squares migration, we consider three levels of preconditioning. First, we correct the order of the normal operator by introducing a left preconditioning, consisting of a fractional 68 3.2. Problem formulation time integration that corresponds to a scaling in the Fourier domain. This first level of preconditioning follows directly from earlier work on migration- amplitude recovery reported by Herrmann et al. (2008a) and Symes (2008). The next level of preconditioning consists of diagonal scaling in the physi- cal domain that compensates for spherical spreading of seismic waves. As a final step, we also include a curvelet-domain scaling as part of the right preconditioning. This step corrects for the remaining amplitude errors that vary spatially as a function of the reflector dip. We conclude by study- ing the performance of these different levels of preconditioning on the SEG AA’ salt model (O’Brien and Gray, 1996; Aminzadeh et al., 1997), using a reverse-time ‘wave-equation’ migration code with optimal checkpointing (Symes, 2007). Because our main interest is to study the performance of our preconditioner, we work on data generated by the linearized Born-scattering operator instead of data generated by the solution of the full nonlinear for- ward model. 3.2 Problem formulation During seismic imaging, the following system of equations needs to be solved Ax ≈ b, (3.1) where b = Ax0 is the multiple-free data with x0 the true reflector model, A the linearized Born-scattering operator, and x the unknown model vector. Here, the symbol ≈ refers to “equality” in the least-squares sense. Con- 69 3.2. Problem formulation trary to many inverse problems, the matrix A, albeit extremely large, is reasonably well behaved and a migrated image can be obtained by applying the adjoint of A to the data vector—i.e., x̃ = A∗b with A∗ the migration operator. The symbol ∗ denotes the adjoint, and x̃ is the estimate (denoted by the˜) of the image. Unfortunately, the output of the above procedure, called migration, pro- duces erroneous results for the amplitudes of the imaged reflectors. To restore these amplitudes, the least-squares solution to Equation 3.1 can be obtained as the solution of the linear system A∗Ax ≈ A∗b, (3.2) with A∗A the normal or Hessian operator. Solutions to this system are not unique and correspond to solutions of the least-squares method—i.e., x̃LS = argmin x 1 2 ‖b−Ax‖22, (3.3) which finds image vectors, x, that after modeling fit the data vector, b. However, the matrix A∗A is not invertible—i.e., it has zero or small singu- lar values that correspond to shadow zones. Because the system is large, we employ an iterative matrix-free solution method, such as LSQR (Paige and Saunders, 1982). By limiting the number of iterations for this method, we control the energy of the solution vector x̃ and we obtain a regularized solution that is equivalent to the solution of a damped least-squares prob- lem (Vogel, 2002). Even though these iterative methods converge relatively 70 3.2. Problem formulation quickly (see e.g. Nemeth et al., 1999; Chavent and Plessix, 1999; Hu et al., 2001; Kuhl and Sacchi, 2003; Yu et al., 2006), the sheer size of the imaging problem calls for a further reduction in the number of iterations. In a perfect world, with infinite computational resources, the ideal pre- conditioning for the system in Equation 3.1 corresponds to AM−1R u ≈ b, x :=M−1R u, (3.4) with the right preconditioning matrix, MR := ( A∗A )1/2, given by the ‘square-root’ of the normal operator. Here, the symbol := refers to ‘defined as’. In this ideal case, migration recovers the image vector x, exactly. Un- fortunately, in practice (albeit some recent exciting progress has been made by Demanet and Ying, 2008, using discrete symbol calculus for smooth sym- bols, a development on which we intend to report in the future) the quantity( A∗A )1/2 cannot be computed and we have to resort to appropriate approx- imations. In this paper, we propose a combination of left and right preconditioning— i.e., we replace Equation 3.1 by bA︷ ︸︸ ︷ M−1L AM −1 R u ≈ bb︷ ︸︸ ︷ M−1L b, x :=M −1 R u, (3.5) with M−1L the left preconditioning matrix. The migrated and least-squares migrated images are given by x̃ = M−1R ũ, with ũ = Â ∗b̂, and by x̃LS = M−1R ũLS , with ũLS = argminu ‖b̂ − Âu‖2, respectively. Our precondi- tioners are derived from the following three observations: (i) under certain 71 3.3. Preconditioning conditions—such as the high-frequency limit, smooth background velocity models, and the absence of turning waves (Stolk, 2000)—the normal oper- ator is in d dimensions a (d − 1)-order pseudodifferential operator (ΨDO, see e.g. recent work by Herrmann et al., 2008a; Symes, 2008, and the refer- ences therein), (ii) migration amplitudes decay with depth due to spherical spreading of seismic body waves, and (iii) zero-order ΨDO’s can be approx- imated by a diagonal scaling in the curvelet domain (see e.g. Herrmann et al., 2008a). These observations allow us to define a series of increasingly more accurate approximations to the ‘square-root’ of the normal operator, leading to better and better preconditioners. Finally, we also argue that using curvelets will add a certain robustness to Gaussian noise and mod- eling errors, an observation substantiated by successful applications of this transform in seismic data processing (Herrmann et al., 2008b; Wang et al., 2008). 3.3 Preconditioning In this section, we introduce different types of preconditioners based on the aforementioned observations. For each preconditioned system, we study the migrated images. Later, we study the convergence of the iterative solver. The examples are computed for the reflectivity and smooth velocity back- ground models plotted in Figure 3.1. To test our preconditioner, data with 324 shots is generated using Equation 3.1. Each shot consists of 176 traces of 6.4s and with a trace interval of 24m. The maximum offset of the data is 4224m. 72 3.3. Preconditioning 3.3.1 Left preconditioning by fractional differentiation As stated before, under aforementioned conditions the normal operator cor- responds to a (d − 1)-order ΨDO. In 2-D image space, this operator corre- sponds to the leading-order behavior of the square root of Laplacian—i.e., the action of ( ∆ )1/2· in the physical domain, or to |ξ|· with ξ the wave vec- tor in the spatial Fourier domain. In data space, this action corresponds to a multiplication by |ω| in the temporal Fourier domain (Herrmann et al., 2008a). We compensate for this action by defining the following left precon- ditioning: M−1L := ∂ −1/2 |t| , (3.6) where ∂−1/2|t| · := F ∗|ω|−1/2F · with F the Fourier transform and F ∗ = F−1 its inverse. We define the level I preconditioner withM−1L as above andM −1 R = I (which implies x = u). Comparison of the migrated images before and after left preconditioning (cf. Figures 3.1(c) and 3.2(a)) shows that the imprint of the Laplacian is removed—i.e., some of the low-frequency content is restored. However, the migrated image still contains dimming of the amplitudes. Note that this left preconditioning corresponds to the solution of the scaled least- squares problem: argminx 1 2‖b̂− Âx‖22 = argminx 12‖M−1L (b−Ax)‖22. 3.3.2 Right preconditioning by scaling in the physical domain To further correct the amplitudes, we propose to apply a scaling to com- pensate for the leading-order amplitude decay. This decay is linear because the reflected waves travel from the source down to the reflector, experi- 73 3.3. Preconditioning encing an amplitude decay proportional to the square-root of the reflector depth (in 2-D). We correct this linear amplitude decay by defining the right preconditioning matrix: M−1R = Dz := diag ( z ) 1 2 , (3.7) where zi = i∆z, i = 1 · · ·nz, with ∆z the depth sample interval and nz the number of samples. Combined with the left preconditioner M−1L , we call this the level II preconditioner. The results for the migrated image in Fig- ure 3.2(b) now show further improvement. However, amplitude variations remain, e.g., along the major horizontal reflector just above 3500m. 3.3.3 Right preconditioning by scaling in the curvelet domain After applying the left preconditioning, the Hessian can be modeled by a zero-order ΨDO whose action corresponds to that of a nonstationary dip filter—i.e., we have ( Ψf ) (x) ' ∫ ξ∈Rd ejξ·xa(x, ξ)f̂(ξ)dξ, (3.8) with Ψ the Hessian of the preconditioned modeling operator and a(x, ξ) a space- and spatial-frequency dependent filter known as the symbol. We use the symbol ' to indicate high-frequency approximation and absence of turning waves. Following ideas that go back to Taylor (1981), a pseudodiffer- ential operator can approximately be diagonalized by linear combinations of 74 3.3. Preconditioning localized oscillatory functions, such as curvelets (Candès et al., 2006), wave atoms (Demanet and Ying, 2007), and local Fourier bases (Meyer, 1992). In this paper, we use recent results by Herrmann et al. (2008a) who use curvelets because they have the additional advantage of being sparse on the model. The action of the ΨDO can, after discretization, be approximated by a scaling in the curvelet domain—i.e., we have the following approximate identity Ψr ≈ C∗D2ΨCr, D2Ψ := diag ( d2 ) , (3.9) which is accurate for a reference vector r close enough to the actual im- age. In this expression, the matrices C and C∗ represent the 2-D discrete curvelet transform (see e.g. Candès et al., 2006) for which the adjoint equals the pseudoinverse—i.e., we have C∗C = I with I the identity matrix. The reciprocal of the curvelet-domain scaling coefficients, d−2, is found by a rem- igrated image-to-image matched-filtering procedure that involves the refer- ence vector, typically derived from a conventional migrated image, and the remigrated reference vector. Hence, the cost of forming Equation 3.9 is ap- proximately one modeling and one migration (for further details refer to Herrmann et al., 2008a,c). As we will show below, these additional costs are well offset by the increase in convergence. By including the above approximation, we define the right precondition- ing matrix as M−1R = DzC ∗D−1Ψ , (3.10) which compensates for the remaining amplitude errors. Because there is no need for high accuracy, Equation 3.9 is a good approximation whose ac- 75 3.4. Convergence of least-squares migration curacy increases with frequency (Herrmann et al., 2008a). Note, however, that this preconditioner by virtue of the underlying assumptions will not help in the removal of nonlocal imaging artifacts. We call the combination of Equations 3.6 and 3.10, the level III preconditioner. The image obtained with this system is plotted in Figure 3.2(c) and shows, as expected, fur- ther improvement in the amplitudes of the migrated image. Despite this improvement, image artifacts and amplitude errors are still present and can be attributed to the approximation in Equation 3.9 and to the fact that migration does not correspond to inversion—i.e., A∗A 6= I . 3.4 Convergence of least-squares migration Even though the different preconditioning operators defined so far lead to im- provements, problems remain in achieving high-fidelity images. As reported in the literature, the image quality can be further improved by replacing migration with least-squares migration where the (preconditioned) scatter- ing matrix is inverted iteratively. After a limited number of iterations, this method inverts the (preconditioned) system of equations approximately. The performance of iterative solvers depends on certain properties of the (preconditioned) matrix A, which include its condition number (ratio of the largest to smallest singular value) and the clustering of the singu- lar values (see e.g. De Roeck, 2002, where these quantities are discussed for a relatively small Kirchhoff-based imaging problem). Because LSQR minimizes the residual during each iteration (cf. Equation 3.3), studying its progress as a function of the number of iterations gives us some way to gauge 76 3.4. Convergence of least-squares migration the performance. For this purpose, we introduce the normalized log-based least-squares residual µk = 20 log ‖Âuk − b̂‖2/‖b̂‖2, with uk the solution of the (preconditioned) system after k iterations and with u0 = 0. Follow- ing De Roeck (2002), we also track the log-based least-squares model-space residual—i.e., νk = 20 log ‖Â∗ ( Âuk − b̂ )‖2/‖Â∗b̂‖2. Note that in practice, these latter residuals are never computed because they are not a by-product of LSQR, and require an additional matrix-vector multiply per iteration. Contrary to data-space residuals, which possibly contain unmodeled com- ponents that may not be in the range of the modeling operator, model-space residuals typically converge to zero. Both quantities are used to empirically establish the performance of the different levels of preconditioning. The re- sults of this exercise are summarized in Figure 3.3, which plots the decay of µk and νk as a function of the number of iterations k. To account for the overhead, plots for the level III preconditioner are offset by one iteration. This is justified because the cost per iteration of evaluating the curvelet transform, and its inverse, is negligible compared to the cost of migration and demigration. As we move from a single left preconditioner, towards left and right pre- conditioners, the data residuals decay faster with a significant improvement obtained by the curvelet-domain scaling. For instance, the residual after 10 iterations for level II preconditioning (fractional integration and depth weighting) is attained by only 5 iterations of level III preconditioning (in- cluding curvelet-domain scaling), whereas the result after 10 iterations is approximately 2 dB better. Including the computational overhead, level III preconditioning gains 8 dB in four iterations. Even though the picture for 77 3.5. Extensions the model-space residual is less clear, there is a similar trend for the level III preconditioning. For instance, after 10 iterations, we have an improvement of approximately 4.5 dB. The improvements in convergence for the preconditioned system are also reflected in the least-squares migrated results included in Figure 3.4. Com- paring these images shows a clear enhancement for the preconditioned sys- tem, plotted in Figure 3.4(b), over the least-squares result obtained without preconditioning. Moreover, juxtaposing the preconditioned least-squares im- age with the solution for the preconditioned migrated image after one itera- tion depicted in Figure 3.2(c) shows a significant enhancement of the overall amplitudes and frequency content (cf. Figure 3.1(a) and 3.4(b)). The bottom line is that each iteration takes 40 and 180 minutes for the modeling and migration on an IBM eServer with 52 processors at 2.2GHz. Aside from one additional modeling and migration, the computation of the diagonal estimation takes approximately 90 minutes on a single CPU. The cost of the 2-D curvelet-transforms part of the preconditioning is less than one minute and is negligible compared to the modeling and migration costs. 3.5 Extensions The preconditioning methodology presented in this letter constitutes a first step towards a concerted effort to formulate seismic imaging and inversion as an optimization problem. This type of formulation allows us to create high-fidelity images and to make progress towards full-waveform inversion. Having access to appropriate preconditioners is instrumental for this purpose 78 3.5. Extensions because it makes the computations tractable. We envisage the following extensions of our work: • generalization to 3-D, which entails a different power for the fractional integration, a different spherical-spreading correction, and a 3-D for- mulation of our curvelet-domain matched filter. Moreover, in 3-D the theory of approximating the Hessian by a ΨDO is less well developed. • replacement of the level II physical-space preconditioning by more accurate source Green’s function illumination corrections (see e.g., Plessix and Mulder, 2004). • inclusion of density variations, or in case of elastic wave propagation, the inclusion of variations in the elastic moduli. This extension re- quires a multi-parameter formulation. • regularization by curvelet-domain sparsity promotion, replacing Equa- tion 3.3 by ũ`1 = argmin u ‖u‖1 subject to ‖Âu− b̂‖2 ≤ (3.11) with a noise-dependent parameter, and x`1 := M −1 R ũ`1 . To ben- efit from our preconditioning, Equation 3.11 requires a solver with Newton-type steps, as opposed to most `1-norm solvers that are based on projected gradients. The advantage of this formulation is that it uses curvelet-domain sparsity, which has proven to be a particularly powerful prior (Wang and Sacchi, 2007; Herrmann et al., 2008b,a; Hen- nenfent et al., 2008). We will report on the solution of Equation 3.11 79 3.6. Conclusions elsewhere. • incorporation of our preconditioner in sparsity-promoting full-waveform inversion. This problem requires the solution of the unconstrained op- timization problem min z 1 2 ‖b̂− F̂[z]‖22 + λ‖z‖1, (3.12) where z is the curvelet representation of the model, λ a Lagrange multiplier balancing the residual energy and `1-norm penalty term, and F̂[z] the nonlinear forward map that links the curvelet-domain model to data. Again, the solution of this optimization problem requires a sophisticated solver that can benefit from our preconditioner. 3.6 Conclusions Because of the size of the seismic imaging problem, preconditioning of least- squares migration is an elusive topic, where traditional approaches from numerical linear algebra have not yet found their way. Lack of direct access to the matrices involved and the cost of evaluating matrix-free implementa- tions of the operators are both to blame. Nonetheless, the first few iterations of the LSQR algorithm of least-squares migration are known to make sig- nificant progress towards the solution. Unfortunately, even for this limited number of iterations, the computational costs are often still prohibitively large for practical problems. The method presented in this paper partly resolves this issue through a combination of left and right preconditioning 80 3.7. Acknowledgments together with a curvelet-domain scaling. Inclusion of the latter proved par- ticularly important because it restores the amplitudes and leads to faster convergence, at a relatively small computational overhead. Aside from this tangible reduction in computational costs of roughly 50%, the use of curvelet frames opens the enticing perspective to use `1-norm regularization to im- prove the quality of images. Preconditioning also plays a pivotal role in making this approach numerically feasible and may extend to a solution of the full-waveform inversion problem with sparsity promotion. Both ap- proaches are justified by ample evidence that curvelets are sparse on the model. 3.7 Acknowledgments The authors would like to thank C. C. Stolk and W. W. Symes for their input on migration-amplitude recovery and for the use of the reverse-time migra- tion code. We also would like to thank Tamas Nemeth and Michael Friedlan- der for insightful discussions and the authors of CurveLab (www.curvelet.org). The examples were prepared with Madagascar (rsf.sf.net), supplemented by SLIMpy operator overloading, developed by S. Ross-Ross. This work was in part financially supported by the NSERC Discovery (22R81254) and CRD Grants DNOISE (334810-05) of F.J.H. and was in part carried out within the SINBAD project with support, secured through ITF, from BG Group, BP, Chevron, ExxonMobil and Shell. We finally thank the anonymous re- viewer and W. W. Symes for their constructive comments and suggestions that greatly improved our paper. 81 3.7. Acknowledgments (a) (b) (c) Figure 3.1: The SEG/EAGE AA′ salt model, (a) reflectivity defined by the high-pass filtered velocity model, (b) smoothed velocity model, (c) the migrated image according to Equation 3.1. This image suffers from dete- riorated amplitudes, especially under the high-velocity salt and for steep reflectors and faults. 82 3.7. Acknowledgments (a) (b) (c) Figure 3.2: Migrated images for different levels of preconditioning, (a) result for left preconditioning (level I, cf. Equation 3.6), (b) result for left-right (including depth-correction) preconditioning (level II, cf. Equation 3.7), (c) the same but now including curvelet-domain scaling (level III, cf. Equation 3.10). 83 3.7. Acknowledgments 0 2 4 6 8 10 Iterations 12 10 8 6 4 2 0 µ k (d B ) None Level I Level II Level III (a) 0 2 4 6 8 10 Iterations 18 16 14 12 10 8 6 4 2 0 ν k (d B ) None Level I Level II Level III (b) Figure 3.3: Residual decays for different levels of preconditioning. The dotted blue lines corresponds to least-squares migration without precondi- tioning, the dash-dotted lines to level I preconditioning, the dashed black lines to level II preconditioning, and the red solid lines to level III precon- ditioning. This is offset by one iteration to account for the overhead, (a) plot for the decay of the data-space normalized residues µk as a function of the number of LSQR iterations, (b) the same but now for the model-space normalized residuals νk. 84 3.7. Acknowledgments (a) (b) Figure 3.4: Least-squares migration without and with preconditioning jux- taposed with the original reflectivity (in light blue), (a) least-squares mi- grated image, (b) least-squares image with level III preconditioning. Notice the improvement in the recovered reflectivity. 85 Bibliography Aminzadeh, F., J. Brac, and T. Kunz, 1997, 3-D Salt and Overthrust Model. SEG/EAGE 3-D Modeling Series, No. 1: Society of Exploration Geophysi- cists, Tulsa. Calvetti, D., 2007, Preconditioned iterative methods for linear discrete ill- posed problems from a Bayesian inversion perspective: Journal of Com- putational and Applied Mathematics, 198, 378–395. Candès, E. J., L. Demanet, D. L. Donoho, and L. Ying, 2006, Fast discrete curvelet transforms: SIAM Multiscale Modeling and Simulation, 5, 861– 899. Chavent, G. and R. E. Plessix, 1999, An optimal true-amplitude least- squares prestack depth-migration operator: Geophysics, 64, 508–515. Claerbout, J. and D. Nichols, 1994, Spectral preconditioning: Technical Report SEP-82, Stanford Exploration Project. De Roeck, Y., 2002, Sparse linear algebra and geophysical migration: A review of direct and iterative methods: Numerical Algorithms, 29, 283 – 322. Demanet, L. and L. Ying, 2007, Wave atoms and sparsity of oscillatory patterns: Journal of Applied and Computational Harmonic Analysis, 23, 368–387. 86 Chapter 3. Bibliography ——–, 2008, Discrete symbol calculus. (Submitted for publication. Available at http://math.stanford.edu/ laurent/papers/DSC.pdf). Guitton, A., 2004, Amplitude and kinematic corrections of migrated images for nonunitary imaging operators: Geophysics, 69, 1017–1024. Hennenfent, G., E. van den Berg, M. P. Friedlander, and F. J. Herrmann, 2008, New insights into one-norm solvers from the Pareto curve: Geo- physics, 73, no. 4, A23–A26. Herrmann, F. J., P. P. Moghaddam, and C. C. Stolk, 2008a, Sparsity- and continuity-promoting seismic imaging with curvelet frames: Journal of Applied and Computational Harmonic Analysis, 24, 150–173. Herrmann, F. J., D. Wang, G. Hennenfent, and P. P. Moghaddam, 2008b, Curvelet-based seismic data processing: a multiscale and nonlinear ap- proach: Geophysics, 73, no. 1, A1–A5. Herrmann, F. J., D. Wang, and D. J. E. Verschuur, 2008c, Adaptive curvelet- domain primary-multiple separation: Geophysics, 73, A17–A21. Hu, J., G. T. Schuster, and P. A. Valasek, 2001, Poststack migration decon- volution: Geophysics, 66, 939–952. Kuhl, H. and M. D. Sacchi, 2003, Least-squares wave-equation migration for AVP/AVA inversion: Geophysics, 68, 262–273. Meyer, Y., 1992, Wavelets and operators: Cambridge University Press. Nemeth, T., C. Wu, and G. T. Schuster, 1999, Least-squares migration of incomplete reflection data: 64, 208–221. O’Brien, M. and S. Gray, 1996, Can we image beneath salt?: The Leading Edge, 15, 17–22. Paige, C. C. and M. A. Saunders, 1982, LSQR: An algorithm for sparse 87 Chapter 3. Bibliography linear equations and sparse least squares: ACM TOMS, 8, 43–71. Plessix, R. and W. Mulder, 2004, Frequency-domain finite difference amplitude-preserving migration: Geophysical Journal International, 157, 975–987. Rickett, J. E., 2003, Illumination-based normalization for wave-equation depth migration: Geophysics, 68, 1371–1379. Stolk, C. C., 2000, Microlocal analysis of a seismic linearized inverse prob- lem: Wave Motion, 32, 267–290. Symes, W. W., 2007, Reverse time migration with optimal checkpointing: Geophysics, 72, SM213–SM221. ——–, 2008, Approximate linearized inversion by optimal scaling of prestack depth migration: Geophysics, 73, R23–R35. Taylor, M., 1981, Pseudodifferential operators: Princeton University Press. Vogel, C., 2002, Computational Methods for Inverse Problems: SIAM. Wang, D., R. Saab, O. Yilmaz, and F. J. Herrmann, 2008, Bayesian wave- field separation by transform-domain sparsity promotion: Geophysics, 73, A33–A38. Wang, J. and M. D. Sacchi, 2007, High-resolution wave-equation amplitude- variation-with-ray-parameter (AVP) imaging with sparseness constraints: Geophysics, 72, S11–S18. Yu, J., J. Hu, G. T. Schuster, and R. Estill, 2006, Prestack migration de- convolution: Geophysics, 71, S53–S62. 88 Chapter 4 Curvelet-based seismic data processing 4.1 Introduction In this letter, we demonstrate that the discrete curvelet transform (Candès et al., 2006a; Hennenfent and Herrmann, 2006b) can be used to reconstruct seismic data from incomplete measurements, to separate primaries and mul- tiples and to restore migration amplitudes. The crux of the method lies in the combination of the curvelet transform, which attains a fast decay for the magnitude-sorted curvelet coefficients, with a sparsity promoting program. By themselves sparsity-promoting programs are not new to the geosciences (Sacchi et al., 1998). However, sparsity promotion with the curvelet transform is new. The curvelet transform’s unparalleled ability to detect wavefront-like events that are locally linear and coherent means it is particularly well suited to seismic data problems. In this paper, we show examples including data regularization (Hennenfent and Herrmann, 2006a, A version of this chapter has been published. Herrmann, F.J., Wang, D., Hennenfent, G. and Moghaddam, P.P. (2008) Curvelet-based seismic data processing: a multiscale and nonlinear approach.Geophysics, 73(1):A1-A5, 2008. 89 4.2. Curvelets 2007a), primary-multiple separation (Herrmann et al., 2007) and migration- amplitude recovery (Herrmann et al., 2008). Application of this formalism to wavefield extrapolation is presented elsewhere (Lin and Herrmann, 2007). 4.2 Curvelets Curvelets are localized ’little plane-waves’ (see Hennenfent and Her- rmann, 2006b, and the on-line ancillary material for an introduction on this topic) that are oscillatory in one direction and smooth in the other direction(s). They are multiscale and multi-directional. Curvelets have an anisotropic shape – they obey the so-called parabolic scaling relationship, yielding a ”width” proportional to square of the ”length” for the support of curvelets in the physical domain. This anisotropic scaling is necessary to detect wavefronts and explains their high compression rates on seismic data and images (Candès et al., 2006a; Herrmann et al., 2008), as long as these datasets can be represented as functions with events on piece-wise twice differentiable curves. Then, the events become linear at the fine scales justifying an approximation by the linearly shaped curvelets. Even seismic data with caustics, pinch-outs, faults or strong amplitude variations fit this model, which amounts to a preservation of the sparsity attained by curvelets. Curvelets represent a specific tiling of the 2-D/3-D frequency domain into strictly localized wedges. Because the directional sampling increases every-other scale doubling, curvelets become more anisotropic at finer scales. Curvelets compose multi-D data according to f = CTCf with C and CT the forward and inverse discrete curvelet transform matrices (defined by the 90 4.3. Common problem formulation by sparsity-promoting inversion fast discrete curvelet transform, FDCT, with wrapping, a type of periodic extenstion, see Candès et al., 2006a; Ying et al., 2005). The symbol T represents the transpose, which is equivalent to the inverse for this choice of curvelet transform. This transform has a moderate redundancy (a factor of roughly 8 in 2-D and 24 in 3-D) and a computational complexity of O(n log n) with n the length of f . Even though CTC = I , with I the identity matrix, the converse is not true, i.e., CCT 6= I . 4.3 Common problem formulation by sparsity-promoting inversion Our solution strategy is built on the premise that seismic data and images have a sparse representation, x0, in the curvelet domain. To exploit this property, our forward model reads y = Ax0 + n (4.1) with y a vector of noisy and possibly incomplete measurements; A the modeling matrix that includes CT ; and n, a zero-centered white Gaussian noise. Because of the redundancy of C and/or the incompleteness of the data, the matrix A can not readily be inverted. However, as long as the data, y, permits a sparse vector, x0, the matrix, A, can be inverted by a sparsity-promoting program (Candès et al., 2006b; Donoho, 2006): 91 4.3. Common problem formulation by sparsity-promoting inversion P : x̃ = argminx ‖x‖1 s.t. ‖Ax− y‖2 ≤ f̃ = ST x̃ (4.2) in which is a noise-dependent tolerance level, ST the inverse transform and f̃ the solution calculated from the vector x̃ (the symbol ˜ denotes a vector obtained by nonlinear optimization) minimizing P. The difference between x̃ and x0 is proportional to the noise level. Nonlinear programs P are not new to seismic data processing as in spiky deconvolution (Taylor et al., 1979; Santosa and Symes, 1986) and Fourier transform-based interpolation (Sacchi et al., 1998). The curvelets’ high com- pression rate makes the nonlinear program P perform well when CT is in- cluded in the modeling operator. Despite its large-scale and nonlinearity, the solution of the convex problem P can be approximated with a limited (< 250) number of iterations of a threshold-based cooling method derived from work by Figueiredo and Nowak (2003); Daubechies et al. (2004); Elad et al. (2005). At each iteration the descent update (x← x+AT (y−Ax)), minimizing the quadratic part of Equation 4.2, is followed by a soft thresh- olding (x ← Tλ(x) with Tλ(x) := sgn(x) · max(0, |x| − |λ|)) for decreasing threshold levels λ. This soft thresholding on the entries of the unknown curvelet vector captures the sparsity and the cooling, which speeds up the algorithm, allows additional coefficients to fit the data. 92 4.4. Seismic data recovery 4.4 Seismic data recovery The reconstruction of seismic wavefields from regularly-sampled data with missing traces is a setting where a curvelet-based method will perform well. As with other transform-based methods, sparsity is used to reconstruct the wavefield by solving P.It is also shown that the recovery performance can be increased when information on the major primary arrivals is included in the modeling operator. 4.4.1 Curvelet-based recovery The reconstruction of seismic wavefields from incomplete data corre- sponds to the inversion of the picking operator R. This operator models missing data by inserting zero traces at source-receiver locations where data is missing passing recorded traces unchanged. The task of the recovery is to undo this operation by filling in the zero traces. Since seismic data is sparse in the curvelet domain, the missing data can be recovered by compounding the picking operator with the curvelet modeling operator, i.e., A := RCT . With this definition for the modeling operator, solving P corresponds to seeking the sparsest curvelet vector whose inverse curvelet transform, fol- lowed by the picking, matches the data at the nonzero traces. Applying the inverse transform (with S := C in P) gives the interpolated data. For details on the conditions that determine successful recovery, refer to Hen- nenfent and Herrmann (2007a,b) and Herrmann and Hennenfent (2007). An example of curvelet-based recovery is presented in Figure 4.1 which shows the results of decimating, and then reconstructing, a seismic dataset. 93 4.4. Seismic data recovery The original shot and receiver spacings were 25m, and 80% of the traces were thrown out at random (see Figure 4.1(b)). Comparing the ’ground truth’ in Figure 4.1(a) with the recovered data in Figure 4.1(c) shows a successful recovery in case the high-frequencies are removed. Aside from sparsity in the curvelet domain, no prior information was used during the recovery, which is quite remarkable. Part of the explanation lies in the curvelet’s ability to locally exploit the 3-D geometry of the data and this suggests why curvelets are successful for complex datasets where other methods may fail. 4.4.2 Focused recovery In practice, additional information on the to-be-recovered wavefield is of- ten available. For instance, one may have access to the predominant primary arrivals or to the velocity model. In that case, the recently introduced fo- cal transform (Berkhout and Verschuur, 2006), which ’deconvolves’ the data with an estimate of the primaries, incorporates this additional information into the recovery process. Application of this primary operator, ∆P, adds a wavefield interaction with the surface, mapping primaries to first-order surface-related multiples (Verschuur and Berkhout, 1997; Herrmann, 2007). Inversion of this operator, strips the data off one interaction with the surface, focusing primary energy to (directional) sources. This focusing correponds to a collapse of the 3-D primary events to an approximate line source which has a sparser representation in the curvelet domain. By compounding the non-adaptive, data-independent, curvelet trans- form with the data-adaptive focal transform, i.e., A := R∆PCT , the re- covery can be improved by solving P. The solution of P now entails 94 4.4. Seismic data recovery (a) (b) (c) (d) Figure 4.1: Comparison between 3-D curvelet-based recovery by sparsity- promoting inversion with and without focusing, (a) fully sampled real SAGA data shot gather, (b) randomly subsampled shot gather from a 3-D data volume with 80% of the traces missing in the receiver and shot directions, (c) curvelet-based recovery, (d) curvelet-based recovery with focusing. No- tice the improvement (denoted by the arrows) from the focusing with the primary operator. 95 4.5. Seismic signal separation the inversion of ∆P, yielding the sparsest set of curvelet coefficients that matches the incomplete data when ’convolved’ with the primaries. Apply- ing the inverse curvelet transform, followed by ’convolution’ with∆P yields the interpolation, i.e. ST :=∆PCT . Comparing the curvelet recovery with the focused curvelet recovery (Figure 4.1(c) and 4.1(d)) shows an overall improvement in the recovered details. 4.5 Seismic signal separation Predictive multiple suppression involves two steps, namely multiple pre- diction and primary-multiple separation. In practice, the second step ap- pears difficult and adaptive least-squares `2-matched-filtering techniques are known to lead to residual multiple energy, high frequency jitter and de- terioration of the primaries (Herrmann et al., 2007). By employing the curvelet’s ability to detect wavefronts with conflicting dips (e.g. caustics), a non-adaptive, independent of the total data, separation scheme can be defined that is robust with respect to moderate errors in the multiple pre- diction. The nonlinear program, P, with y defined by the total data, can be adapted to separate multiples from primaries by replacing the `1 norm by a weighted `1 norm, i.e., ‖x‖1 7→ ‖x‖1,w = ∑ µ |wµxµ| with µ running over all curvelets and w a vector with positive weights. By defining these weights proportional to the magnitude of the curvelet coefficients of the 2-D SRME-predicted multiples, the solution of P with A := CT removes mul- tiples. Primaries and multiples naturally separate in the curvelet domain and the weighting further promotes this separation while solving P. The 96 4.5. Seismic signal separation weights that are fixed during the optimization penalize the entries in the curvelet vector for which the predicted multiples are significant. The em- phasis on the weights versus the data misfit (the proportionality constant) is user defined. The estimate for the primaries is obtained by inverse curvelet transforming the curvelet vector that minimizes P for the weighted `1 norm (A = ST := CT ). Figure 4.2 shows an example of 3-D curvelet-based primary-multiple separation of a North Sea dataset with the weights set according to the curvelet-domain magnitudes of the SRME-predicted multiples multiplied by 1.25. Comparison between the estimates for the primaries from adaptive subtraction by `2-matched filtering (Verschuur and Berkhout, 1997) and from our nonlinear and non-adaptive curvelet-based separation shows an improvement in (i) the elimination of the focused multiple energy below shot location 1000m, induced by out-of-plane scattering due to small 3-D variations in the multiple-generating reflectors and (ii) an overall improved continuity and noise reduction. This example demonstrates that the multi- scale and multi-angular curvelet domain can be used to separate primaries and multiples given an inaccurate prediction for the multiples. However, the separation goes at the expense of a moderate loss of primary energy which compares favorably compared to the loss associated with `2-matched filtering (see also Herrmann et al., 2007). 97 4.5. Seismic signal separation (a) (b) (c) (d) Figure 4.2: 3-D Primary-multiple separation with P for the SAGA dataset, (a) near-offset section including multiples, (b) the SRME-predicted multi- ples, (c) the estimated primaries according to `2-matched filtering, (d) the estimated primaries obtained with P. Notice the improvement, in areas with small 3-D effects (ellipsoid) and residual multiples. 98 4.6. Migration-amplitude recovery 4.6 Migration-amplitude recovery Restoring migration amplitudes is another area where curvelets can be shown to play an important role. In this application, the purpose is to replace computationally expensive amplitude recovery methods, such as least-squares migration (Nemeth et al., 1999; Kuhl and Sacchi, 2003), by an amplitude scaling (Guitton, 2004). This scaling can be calculated from a demigrated-migrated reference vector close to the actual reflectivity. In order to exploit curvelet sparsity, we propose to scale in the curvelet domain. This choice seems natural because migrated images suffer from spatially varying and dip-dependent amplitude deterioration that can be accommodated by curvelets. The advantages of this approach are manifold and include (i) a correct handling of reflectors with conflicting dips and (ii) a stable curvelet sparsity-promoting inversion of the diagonal that restores the amplitudes and removes the clutter by exploiting curvelet sparsity on the model. The method is based on the approximate identity: KTKr ≈ CTDrCr with K and KT the demigration, migration operators and Dr a reference- model specific scaling (Herrmann et al., 2008). By defining the modeling matrix as A := CT √ Dr, P can be used to recover the migration am- plitudes from the migrated image. Possible spurious side-band effects and erroneously detected curvelets (Candès and Guo, 2002) are removed by sup- plementing the `1 norm in P with an anisotropic diffusion norm (Fehmers and Höcker, 2003). This norm enhances the continuity along the imaged reflectors and removes spurious artifacts. 99 4.7. Discussion and conclusions Results for the SEG AA’ dataset (O’Brien and Gray, 1996; Aminzadeh et al., 1997) are summarized in Figure 4.3. These results are obtained with a reverse-time ’wave-equation’ finite-difference migration code. To illustrate the recovery performance, idealized seismic data is generated by demigra- tion, followed by adding white Gaussian noise, yielding a signal-to-noise ratio (SNR) of only 3 dB. This data is subsequently migrated and used as input. Despite the poor SNR, the image in Figure 4.3(a) contains most reflectors, which can be explained by the redundancy of the data, the migration op- erator’s sophistication (diffractions at the bottom of the salt are handled correctly) and the perfect ’match’ between the demigration and migration operators. However, the noise gives rise to clutter and there is dimming of the amplitudes, in particular for steep dips under the salt. Nonlinear recov- ery removes most of this clutter and more importantly the amplitudes for the sub-salt steep-dipping events are mostly restored. This idealized exam- ple shows how curvelets can be used to recover the image amplitudes. As long as the background velocity model is sufficiently smooth and the reflec- tivity sufficiently sparse, this recovery method can be expected to perform well even for more complex images. 4.7 Discussion and conclusions The presented examples show that problems in data acquisition and imaging can be solved with a common problem formulation during which sparsity in the curvelet domain is promoted. For curved wavefront-like fea- tures that oscillate in one direction and that are smooth in the other di- 100 4.7. Discussion and conclusions (a) (b) Figure 4.3: Image amplitude recovery for a migrated image calculated from noisy data (SNR 3dB), (a) image with clutter, (b) image after nonlinear recovery. The clearly visible non-stationary noise in (a) is mostly removed during the recovery while the amplitudes are also restored. Steeply dipping reflectors (denoted by the arrows) under the salt are also well recovered. 101 4.8. Acknowledgments rection(s), curvelets attain high compression rates while other transforms do not necessarily achieve sparsity for these geometries. Seismic images of sedimentary basins and seismic wave arrivals in the data both behave in this fashion, so that curvelets are particularly valuable for compression. It is this compression that underlies the success of our sparsity promoting formulation. First, we showed on real data that missing data can be re- covered by solving a nonlinear optimization problem where the data misfit and the `1-norm on the curvelet coefficients are simultaneously minimized. This recovery is improved further with a combined curvelet-focal transform. Sparsity also proved essential during the primary-multiple separation. In this case, it leads to a form of decorrelation of primaries and multiples, re- ducing the probability of having large overlapping curvelet entries between these different events. Finally, the sparsity of curvelets on the image itself was exploited to recover the migration amplitudes of the synthetic subsalt imaging example. Through these three examples, the successful application of curvelets, enhanced with sparsity-promoting inversion, opens new per- spectives on seismic data processing and imaging. The ability of curvelets to detect wavefront-like features is key to our success and opens an exciting new outlook towards future developments in exploration seismology. 4.8 Acknowledgments The authors would like to thank D.J. Verschuur and C. Stolk for their input in the primary-multiple separation and migration-amplitude recovery. We also would like to thank the authors of CurveLab (www.curvelet.org) 102 4.8. Acknowledgments and W. Symes for his reverse-time migration code. The examples were prepared with Madagascar (rsf.sf.net), supplemented by SLIMpy operator overloading, developed by S. Ross Ross. Norsk Hydro is thanked for the field dataset. M. O’Brien, S. Gray and J. Dellinger are thanked for the SEG AA’ data. This work was in part financially supported by the NSERC Discovery (22R81254) and CRD Grants DNOISE (334810-05) of F.J.H. and was carried out as part of the SINBAD project with support, secured through ITF, from BG Group, BP, Chevron, ExxonMobil and Shell. Finally, the authors would like to thank the anonymous reviewers whose constructive comments helped improve this letter. 103 Bibliography Aminzadeh, F., J. Brac, and T. Kunz, 1997, 3-D Salt and Overthrust Model. SEG/EAGE 3-D Modeling Series, No. 1: Society of Exploration Geophysi- cists, Tulsa. Berkhout, A. J. and D. J. Verschuur, 2006, Focal transformation, an imaging concept for signal restoration and noise removal: Geophysics, 71, A55– A59. Candès, E. J., L. Demanet, D. L. Donoho, and L. Ying, 2006a, Fast discrete curvelet transforms: SIAM Multiscale Modeling and Simulation, 5, 861– 899. Candès, E. J. and F. Guo, 2002, New multiscale transforms, minimum total variation synthesis: Applications to edge-preserving image reconstruction: Signal Processing, 82, 1519–1543. Candès, E. J., J. K. Romberg, and T. Tao, 2006b, Stable signal recovery from incomplete and inaccurate measurements: Communications on Pure and Applied Mathematics, 59, 1207–1223. Daubechies, I., M. Defrise, and C. De Mol, 2004, An iterative thresholding algorithm for linear inverse problems with a sparsity constraints: Com- munications on Pure and Applied Mathematics, 57, 1413–1457. Donoho, D. L., 2006, Compressed sensing: IEEE Transactions on Informa- 104 Chapter 4. Bibliography tion Theory, 52, 1289–1306. Elad, M., J. L. Starck, P. Querre, and D. L. Donoho, 2005, Simultaneous Cartoon and Texture Image Inpainting using Morphological Component Analysis (MCA): Journal of Applied and Computational Harmonic Anal- ysis, 19, 340–358. Fehmers, G. C. and C. F. W. Höcker, 2003, Fast structural interpretation with structure-oriented filtering: Geophysics, 68, 1286–1293. Figueiredo, M. and R. Nowak, 2003, An EM algorithm for wavelet-based image restoration: IEEE Transactions on Image Processing, 12, 906–916. Guitton, A., 2004, Amplitude and kinematic corrections of migrated images for nonunitary imaging operators: Geophysics, 69, 1017–1024. Hennenfent, G. and F. J. Herrmann, 2006a, Application of stable signal recovery to seismic interpolation: Presented at the SEG International Exposition and 76th Annual Meeting. ——–, 2006b, Seismic denoising with non-uniformly sampled curvelets: IEEE Computing in Science and Engineering, 8, 16–25. ——–, 2007a, Irregular sampling: from aliasing to noise: Presented at the EAGE 69th Conference & Exhibition. ——–, 2007b, Random sampling: new insights into the reconstruction of coarsely-sampled wavefields: Presented at the SEG International Exposi- tion and 77th Annual Meeting. Herrmann, F. J., 2007, Surface related multiple prediction from incomplete data: Presented at the EAGE 69th Conference & Exhibition. Herrmann, F. J., U. Boeniger, and D. J. Verschuur, 2007, Nonlinear primary- multiple separation with directional curvelet frames: Geophysical Journal 105 Chapter 4. Bibliography International, 170, 781–799. Herrmann, F. J. and G. Hennenfent, 2007, Non-parametric seismic data recovery with curvelet frames: Technical report, UBC Earth and Ocean Sciences Department. (TR-2007-1). Herrmann, F. J., P. P. Moghaddam, and C. Stolk, 2008, Sparsity- and continuity-promoting seismic imaging with curvelet frames: Applied and Computational Harmonic Analysis. Kuhl, H. and M. D. Sacchi, 2003, Least-squares wave-equation migration for AVP/AVA inversion: Geophysics, 68, 262–273. Lin, T. and F. J. Herrmann, 2007, Compressed wavefield extrapolation: Geophysics. Nemeth, T., C. Wu, and G. T. Schuster, 1999, Least-squares migration of incomplete reflection data: Geophysics, 64, 208–221. O’Brien, M. and S. Gray, 1996, Can we image beneath salt?: The Leading Edge, 15, 17–22. Sacchi, M. D., T. J. Ulrych, and C. Walker, 1998, Interpolation and extrap- olation using a high-resolution discrete Fourier transform: IEEE Trans- actions on Signal Processing, 46, 31–38. Santosa, F. and W. Symes, 1986, Linear inversion of band-limited reflection seismogram: SIAM Journal on Scientific and Statistical Computing, 7, 1307–1330. Taylor, H. L., S. C. Banks, and J. F. McCoy, 1979, Deconvolution with the `1 norm: Geophysics, 44, 39–52. Verschuur, D. J. and A. J. Berkhout, 1997, Estimation of multiple scattering by iterative inversion, part II: practical aspects and examples: Geophysics, 106 Chapter 4. Bibliography 62, 1596–1611. Ying, L., L. Demanet, and E. J. Candés, 2005, 3-D discrete curvelet trans- form: Wavelets XI, Expanded Abstracts, 591413, SPIE. 107 Chapter 5 True amplitude depth migration using curvelets 5.1 Introduction Despite the fact that modern depth migrations can accurately position re- flecting surfaces in the Earth subsurface, they often do not correctly recover high-fidelity amplitude information. This is true even when accurate knowl- edge of the velocity model is available. Many approaches have been proposed to solve the “true-amplitude” mi- gration problem and they generally fall into three main categories. First, there are approaches that correct for the amplitude “during migration” by applying imaging conditions that take reflector amplitudes into account (see e.g. J. C. Costa and Novais, 2009; Chattopadhyay and McMechan, 2008; J. Schleicher and Novais, 2008; Y. Zhang and Zhang, 2007). The second category involves linearized inversion where the linearized scattering operator, i.e. the adjoint of migration (see e.g. Guitton, 2004; Claerbout, 1985; Gray, 1997; Symes, 2007), is inverted using iterative Lanc- A version of this chapter will be submitted for publication. Moghaddam, P.P., Her- rmann F. J. and Shahidi, R. (2010) True amplitude depth migration using curvelets. 108 5.1. Introduction zos methods (Herrmann et al. (2009)). In this category, a variety of algo- rithms for true amplitude migration have been introduced which iteratively update the migrated image to minimize the misfit between observed and the simulated data (see e.g. Tarantola, 1987; Nemeth et al., 1999; Kuehl and Sacchi, 2003; Mulder and Plessix, 2004; Herrmann et al., 2009). Third, there are the so called image-to-image scaling methods, where the normal operator (i.e. migration followed by the scattering operator) is approximated through a transform-domain scaling method. The scaling is done by comparing the migrated image (with some possible pre/post- processings) with a second image obtained by remigration of the image (i.e. after applying the normal operator on the migrated image). Typically, this diagonal scaling is performed in some transform domain ( Fourier, curvelets, physical, etc, see e.g. Rickett, 2003; Plessix and Mulder, 2004; Guitton, 2004; Symes, 2008a; Herrmann et al., 2008). Our method falls in the third category, and we address the amplitude re- covery problem by exploiting the recently developed curvelet frames. These frame expansions compress seismic images (see e.g. Hennenfent and Her- rmann, 2006; E. J. Cand‘es and Ying, 2006) and consist of a collection of frame elements ‘curvelets’ that are invariant under the normal operator (Her- rmann et al. (2008, 2009)). These two properties allow for the development of an approach where the normal operator is nonlinearly inverted, using the eigen-like behavior of curvelets, in combination with the ability of curvelets to handle conflicting dips and sparsely represent the seismic images. The main contributions of this chapter are three-fold. First, we replace the linear least-squares formulation for the estimation of the curvelet-domain 109 5.2. Outline of the chapter coefficients (Herrmann et al. (2008)) by a nonlinear least-squares formula- tion that approximates the symmetric positive definite normal operator by imposing positivity on the scaling coefficients. Second, we apply this method to correct the amplitudes of images that suffer from migration noise caused by small number of shots, as well as noise in the data. Third, we compare two sparsity-promoting methods for recovery, a soft thresholding-correction technique and a large scale sparsity solver. Our method is tested with reverse-time ‘wave equation’ migration code simulating the acoustic wave equation on different synthetic models. 5.2 Outline of the chapter In the first section, we derive our formulation and propose two methods for amplitude recovery. Next, we derive the diagonal estimation for the normal operator in the curvelet domain. The chapter ends with an example of each method on the SEG salt model linearized Born dataset and comparison of the proposed methods followed by conclusion. 5.3 Problem Formulation By assuming our data consist of primaries only (i.e., the surface related and all the internal multiples are assumed to be removed before migration), the data can be expressed as, d ≈ Km+ n, (5.1) 110 5.3. Problem Formulation with d data, K the linearized Born scattering operator (also known as the demigration operator), m the perturbation in the velocity model (i.e., δv v3 with v as velocity, see William W. Symes and Gao (2004); Symes (2008b) ) and n zero-centered Gaussian noise with variance σ2n. Theoretically, migra- tion is the adjoint of the scattering operator. After applying the migration to the data, we obtain the image, y = KTd (5.2) = Ψm+KTn (5.3) where y is the migrated data, KT is the migration operator and the Ψ is the normal operator defined byΨ = KTK. In order to findm in above equation, we need to invert the normal operator (Ψ). To solve the above equation, we follow recent results by Herrmann et al. (2008, 2009) and approximate the normal operator by a curvelet-domain scaling, i.e.,we have, Ψm ' CTWCm, (5.4) with W a diagonal matrix. This diagonal approximation involves applying the curvelet transform C, followed by a scaling with the positive diagonal matrix W and the inverse curvelet transform back to the physical domain by CT . [Note that the curvelets are a tight frame (E. J. Cand‘es and Ying, 2006) so the adjoint of curvelet transform (CT ) is equal to its pseudo-inverse 111 5.3. Problem Formulation (i.e., CTC = I)]. Inserting this approximation into Equation 5.2 leads to, y ' CTWCm+ e, (5.5) with e = KTn colored noise with covariance, Cee = E(eeT ) = E(KTnnTK) = KTE(nnT )K = σ2nK TK, (5.6) with E(.) the statistical expectation and Cnn = E(nnT ) = σ2nI the covari- ance of the white Gaussian noise. To exploit curvelet-domain sparsity on the recovered model, we rewrite Equation 5.5 as, y = CTWx+ e (5.7) where x = Cm. In the context of stable signal recovery (see e.g. Candès et al., 2006; Daubechies et al., 2005), x can be found by solving the following optimization problem, P1 : x̃ = argminx ||x||1 = ∑ µ∈M |xµ| subject to ||C − 1 2 ee (y −CTWx)||2 ≤ υ m̃ = CT x̃, (5.8) where υ is the noise dependent tolerance, .̃ stands for the estimated value and µ is the index set of curvelets. The term C − 1 2 ee in front of misfit acts as the whitening operator for the data misfit (Tarantola (1987)). During the optimization, the vector x is found by minimizing the `1 norm on the curvelet representation for the model and the `2 norm for data misfit. Equation 5.8 112 5.3. Problem Formulation allows us to exploit curvelet-domain sparsity subject to fitting the migrated image within a tolerance by solving the sparsity-promoting program. Inserting Equation 5.4, the approximation for the normal operator into Equation 5.6 for the noise covariance yields, Cee = σ2nK TK ' σ2nCTWC. (5.9) Assuming the curvelet transform is nearly orthonormal, we can approximate the action of C − 1 2 ee in Equation 5.8 by, C − 1 2 ee ' σ−1n CTW− 1 2C. (5.10) With this approximation, Equation 5.8 becomes, P : x̃ = argminx ||x||1 subject to ||b−C TW 1 2x||2 ≤ m̃ = CT x̃, (5.11) with b = CTW− 1 2Cy and = σnυ. We propose two alternative methods for solving above problem, one is a Basis Pursuit De-Noising problem (BPDN) and other one is a soft-thresholding solution. BPDN proposes a solution to problem P equivalent to a weighted `1 problem (by changing the variable W 1 2x 7→ x) yielding m̃ = CTW− 12 x̃, BPDN : x̃ = argmin x ||x|| 1,W− 1 2 = ∑ i |xiw−1/2i | subject to ||b−CTx||2 ≤ (5.12) We solve optimization BPDN with a projected gradient method (SPGL1), 113 5.4. Diagonal Approximation of the Normal Operator recently introduced by (van den Berg and Friedlander, 2008). An alternative way can be evaluating P in unconstrained form. This is equivalent to a quadratic program yielding m̃ = CT x̃, QP : x̃ = argminx 12 ||b−CTx||22 + λ()||x||1,W− 12 (5.13) where λ() is a constant controlled by the noise level in the data. An ap- proximate solution for the optimization QP can be found with a weighted soft-thresholding method, given by, m̃ = CT sign(Cb)max(0, |(Cb)w− 12 | − λ()), (5.14) where is the element-wise product, w = diagW is a vectorized form of the diagonal elements for the matrix W, sign(·) is the element-wise sign function and | · | the absolute value function. In the examples section, we examine both methods in term of quality of their output. 5.4 Diagonal Approximation of the Normal Operator In the previous section, we stated that we can diagonally approximate the normal operator in the curvelet domain. This approximation (c.f., Equa- tion 5.4), replaces the construction of the normal operator, which is com- putationally expensive, with a curvelet domain diagonal scaling, which is computationally cheap. In this section, we provide the theoretical underpin- ning of such approximation by introducing the curvelets and their invariance 114 5.5. Curvelets and their invariance under normal operator bound under the action of normal operator. We then propose a method to approximate the normal operator with a diagonal scaling in curvelet domain. 5.5 Curvelets and their invariance under normal operator A 2D curvelet φµ is defined by its index µ = (j, k, θ) with j = 0, 1, 2, ... con- sisting of its scale subindex, θ = 0, 1, ..., 2bj/2c − 1, its orientation subindex (bxc is the lower integer part of x ), and k = (kx, kz) the location in the wavenumber domain subindex which are scale and angle dependent. Fig- ure 5.1(a) shows an example of different curvelets with different scales and orientations and Figure 5.1(b) shows the corresponding Fourier spectrum. An important feature of curvelets is that the action of the normal oper- ator (i.e., Ψ = KTK ) on a curvelet φµ corresponds to its multiplication by a positive scalar. In Herrmann et al. (2008), we proved a theorem bound- ing the error with the normal operator diagonal approximation in curvelet domain, i.e., we showed, ||Ψφµ −CTWCφµ||2 ≤ β2−j/2, (5.15) with W a positive diagonal matrix as defined before and β a constant. According to this theorem, the approximation error decays with scale (i.e., j), which is consistent with the high-frequency asymptotic behavior that underlies the diagonal approximation. Shahidi and Herrmann (2009a) shows that the constant β is of moderate size. 115 5.5. Curvelets and their invariance under normal operator k1 0-0.25-0.50 0.25 0.50 -0.50 k 2 -0.25 0 0.25 0.500 100 200 300 400 500 0 100 200 300 400 500 Samples S am pl es Figure 5.1: Spatial and frequency representation of curvelets, (a) four dif- ferent curvelets in the spatial domain at three different scales, (b) dyadic partitioning in the frequency domain, where each wedge corresponds to the frequency support of a curvelet in the spatial domain. This figure illustrates the microlocal correspondence between curvelets in the physical and Fourier domain. Curvelets are characterized by rapid decay in the physical space and of compact support in the Fourier space. Notice the correspondence between the orientations of curvelets in the two domains. The 90o rotation is a property of the Fourier transform. Courtesy of Herrmann et al. (2008). 116 5.6. Diagonal approximation The normal operator is pseudodifferential when no multipathing occurs (Beylkin, 1985; Rakesh, 1988; Ten-Kroode et al., 1998), and its order in space dimension n is n − 1 (Stolk, 2000). The inequality in Equation 5.15 holds when the order of normal operator is zero (Herrmann et al., 2008). Ten-Kroode et al. (1998); Stolk (2000) show that in 2D applying an operator of order −1/2 to both migration and modeling operators makes the operator of zero order. As it is shown in Herrmann et al. (2009), this is equivalent to fractional integration of the source wavelet in both the modeling and migra- tion operators. Mathematically, this can be written as ∂−1/2t · = F−∗|ω|−1/2· with F−∗ the inverse Fourier transform. Equation 5.15 states that curvelets are similar to eigenvectors of the normal operator. Therefore, we propose a quasi-eigenvalue decomposition for the normal operator with curvelets as quasi-eigenvectors. This type of approach is similar to the Wavelet-Vaguellette Decomposition (WVD) proposed by (Donoho, 1995; Herrmann et al., 2008). 5.6 Diagonal approximation In Herrmann et al. (2008), we derive an approximation of the normal oper- ator by remigrating the migrated image again (i.e., mremig = KTKmmig). Subsequently, find a diagonal approximation for the normal operator using these image pairs. Mathematically, mremig = (KTK)mmig ' CTWCmmig (5.16) ' CTdiag(v)w, (5.17) 117 5.6. Diagonal approximation with v = Cmmig and w representing the diagonal elements of W (i.e., W = diag(w)). Equation 5.16 aims to find a curvelet-domain diagonal scaling vector (i.e., w in Equation 5.16). Because the curvelet transform is redundant (i.e., it has more rows than columns), inverting for w in Equation 5.16 needs solv- ing an underdetermined system of equations yielding non-unique solutions. To find a unique solution to this estimation problem, we exploit additional properties of the normal operator. First, the normal operator is positive def- inite, yielding positive entries for the scaling vector w. Second, the normal operator is elliptic Pseudo-differential (Stolk and De-Hoop, 2006; Shahidi and Herrmann, 2009a) and is invertible when the symbols of this operator are smooth in both physical and Fourier space, when the background veloc- ity model is smooth (Stolk and De-Hoop, 2006; Herrmann et al., 2008). We use these two properties to formulate the diagonal estimation in terms of a nonlinear least-squares problem, which yields positivity of the entries of the diagonal and smoothness amongst neighboring curvelet coefficients in space and angle domain. Compared to our earlier work (Herrmann et al., 2008), the positivity assumption is new in the context of imaging (see related work on application of this method for multiple removal, Herrmann et al. (2007) ) and the diagonal estimated by following, Pd : z̃ = argminz J(z) = 1 2 ||mmig −CTdiag(v)ez||2 + κ||Lz||2, w̃ = (diag)(w̃) with w̃ = ez̃, (5.18) where w is replaced by ez to ensure positivity of the estimated diagonal 118 5.7. Practical Workflow elements (W). In above equation, phase-space smoothness is promoted by minimizing `2 norm of the coefficients after applying the sharpening oper- ator, L = [Dx Dz Dθ]. This operator penalizes fluctuations amongst neighboring coefficients in ez, i.e., coefficients that are close in phase space both in position and angle (for more details on this topic see Shahidi and Herrmann (2009b)). The matrices Dx,z contain first-order differences at each scale in the x, z directions, and Dθ the first-order difference in the θ direction. Because the curvelet grid differs for each scale and angle, this sharpening operator is implemented for each curvelet wedge separately. κ is a smoothness parameter that control the smoothness of the scaling versus data misfit. In this paper, we select this parameter by exhaustive search for automatic parameter selection (Shahidi and Herrmann (2009a)). Equation 5.18 is convex and can be minimized by an efficient numerical optimization method. We use Limited-memory BFGS (see e.g. Nocedal and Wright, 1999), using following gradient, ∇zJ(z) = ez diag(v)C(CTdiag(v)ez −mmig) + 2κLTLz. (5.19) 5.7 Practical Workflow Our algorithm consists of two main steps, namely the calculation of the curvelet scaling coefficients via an image-to-remigrated image matched fil- tering and the subsequent estimation of amplitude-corrected, denoised image by sparsity promotion. 119 5.7. Practical Workflow In order to apply the curvelet-domain match filter, we require a curvelet implementation (E. Candes (2005)). Our algorithm also requires access to the Born modeling operator and its adjoint, the migration operator. For this purpose, we use Symes (2007) reverse-time migration with optimal check- pointing and its linearized modeling. We make the normal operator zero order by applying a zero-phase half- time integration to the source wavelet input to both operators (Ten-Kroode et al., 1998; Symes, 2008a). Our algorithm consists of the following steps: 1. Perform migration and depth correction–i.e., d 7→ DZKTd = mmig, with mmig depth corrected migrated image and DZ = diag(z) where zi = i4z, for i = 1 : nz, 4z is the depth grid spacing and nz is the number of samples in depth (Herrmann et al. (2009)). The reason for depth correction is to amplify the amplitudes of deep events in the migrated image. This ensures that after applying the normal operator, the amplitude of the deep events are partially calibrated so it can be used to extract information from the normal operator via estimation of its diagonal in curvelet domain. 2. Remigrate and resimulate the migrated image–i.e.,KTKm̃mig =mremig. 3. Find the curvelet domain diagonal scalingW for which m̃mig ' CTWCmremig by solving the optimization problem Pd (cf., Equation 5.18). 4. Construct the original image by solving the minimization problem QP with weighted soft-thresholding (Equation 5.14) or the optimization problem BPDN using a solver for large-scale sparse reconstruction. 120 5.8. Examples 5.8 Examples In this section, we apply our amplitude recovery methods on the (noisy) linearized Born dataset with noise using SEG-AA salt model (Aminzadeh et al. (1997); O’Brien and Gray (1996)). The SEGAA model shown in Figure 5.2(a) is 5.28km in depth and 16.8km in offset with grid resolution of 12m. The 5000 time-step linearized Born modeling takes about 24 min with 25 CPUs on a MPI cluster, while the migration takes about 90 min. In this section, we compare the soft-thresholding and the BPDN methods described in this paper in terms of output quality and calculation speed for both no noise in data (migration noise only) and when noise is added to data. We also analyze the accuracy of the estimation by comparing the recovered image traces with the reflectivity. 5.8.1 Noise-free case Since the removal of the migration noise is our primary interest, we conduct an imaging experiment with a small number of shots. This dimensionality reduction results in cheaper acquisition and faster computation but goes at the expense of introducing the migration noise. This experiment serves as an illustration on how well our method performs in both the removal of migration noise and restoration of the amplitudes. We use a land-acquisition scenario with 50 shots between 50m and 16.5km at a 336m spacing and 680 receivers per shot at a 24m interval with offsets between 200m and 16.5km. To generate the linearized Born dataset, we smooth this velocity model as shown in Figure 5.2(b) and use the difference 121 5.8. Examples (a) (b) (c) Figure 5.2: Conflicting dips velocity model and reflectivity (a) velocity model, (b) smooth velocity model used to generate our example, (c) re- flectivity generated by subtracting smooth velocity model from original one 122 5.8. Examples between the smooth and the original velocity as the reflectivity shown in Figure 5.2(c). Figure 5.3 summarizes our approximation of the normal operator with the curvelet-domain scaling. We chose the reflectivity in Figure 5.2(c) and applied the scattering operator to generate linearized Born dataset (i.e., d = Km). We used the smoothed velocity model shown in Figure 5.2(b) as background velocity for the scattering operator. The depth corrected migrated result is included in Figure 5.3(a). After another modeling and migration of this image, remigrated image is shown in Figure 5.3(b). These two images serve as input to our curvelet scaling coefficient estimation. Ap- plying this curvelet-domain scaling to the reference image yields a good ap- proximation to the action of normal operator as can be seen in Figure 5.3(c). The relative approximation error in this case is 3.71%. To illustrate the importance of imposing curvelet domain smoothness during the estimation of diagonal scaling, we compare the results of two experiments, one without κ = 0 in Equation 5.19, and one with a smoothing regularization (κ = 0.01 found empirically and used to generate the example in Figure 5.3. Figure 5.4 shows the plot of coefficients for each wedge of the reference vector (Figure 5.4(a)), the diagonal estimation without smoothness (Figure 5.4(b)) and with the smoothness (Figure 5.4(c)) constraints. These figures show curvelet decomposition of a shot gather at different frequency band (scale) and angles (dip). Five scales are used for the decomposition. The centre (coarsest scale) shows the DC and low frequency components of the shot gather. The second coarsest scale has 8 angles. The number of angles is doubled to 16 at the third and fourth scale and doubled again to 123 5.8. Examples (a) (b) (c) Figure 5.3: Diagonal approximation of normal operator, (a) reference image (depth corrected migrated image), (b) normal operator on the reference image (i.e., Ψr) (c) approximate normal operator on the reference image (i.e., CTWCr), approximation error is %3.71 124 5.8. Examples 32 for the fifth scale. Note the portions of shot gather captured at various angles and scales. As one can see, the curvelet coefficients for the non zero smoothing regularization κ = .01 (Figure 5.4(c)) is smooth over different angles and locations. Soft-thresholding solution We investigate the effect of the threshold λ on the soft-thresholding solution for the optimization problem QP. Figure 5.5 shows the result of solving the optimization problem QP with soft thresholding for different values of the threshold λ. The solution of soft thresholding shown in Figure 5.5(a)is for when λ = 0 . This solution is similar to approximately inverting the normal operator in the curvelet domain and applying it to the migrated image (i.e., in Equation 5.11 m̃ = CTW− 1 2Cb). Figure 5.5(b) shows the result for λ = σb and Figure 5.5(c) for λ = 2σb). We clearly see from the figures that by increasing λ the incoherent noise removed from the image at the expense of dimming the amplitude of the bottom reflectors. BPDN solution Here we investigate the effect of the tolerance on the solution for the optimization problemBPDN for noise-free data. Figure 5.6 show the results for solving the optimization problem BPDN using SPGL1 solver (Berg and Friedlander, 2007; van den Berg and Friedlander, 2008) for different value of the tolerance . The solution of SPGL1 shown in Figure 5.6(a) is for when is small ( = σb with σb is the standard deviation of b in BPDN). This solution is similar to Figure 5.5(a) and corresponds to applying the 125 5.8. Examples scales Angles (a) scales Angles (b) scales Angles (c) Figure 5.4: Curvelet-domain representation of the curvelet vector obtained from (a) reference vector, (b) scaling coefficients without smoothing con- straints, (c), scaling coefficients with smoothing constraint. The different sub-images represent the curvelet coefficients at different scales (coarsest in the center) and different angles. 126 5.8. Examples (a) (b) (c) Figure 5.5: Examples of the solution for optimization problem QP with soft-thresholding, (a) λ = 0, (b) λ = σb (c) λ = 2σb 127 5.8. Examples inverse of the diagonal approximation to b. Figure 5.6(b) shows the result for larger ( = 50σb) and Figure 5.6(c) is the solution when the is very high ( = 100σb). As can be seen from the figures, when is small, the image contains many artifacts that increases with depth and when the is large, these artifacts removed at the expense of dimming the bottom reflectors. The best choice of is shown in Figure 5.6(b) which is the result of testing a range of different value in the solver. In this image, the deep artifacts are visibly removed and the bottom reflectors are enhanced. Amplitude Analysis To investigate the performance and accuracy of our amplitude recovery method, we examine a number of traces in the recovered images and com- pare them to the original reflectivity. Figure 5.7 shows the location of the traces. To examine the performance of our method in the presence of a salt body, we choose a number of traces inside and outside the salt body. The offset location of traces are (3480, 5160, 5880, 8640, 10080, and 12720) m, respectively. Figure 5.8 shows the amplitude of the traces along the off- set lines (3480, 5160, and 5880) m, respectively. Except small amount of residual noise in the recovered images, all the strong phases well recovered. Early arrivals are generally well recovered; however, with increasing depth the amount of noise increases. Figure 5.9 shows the amplitude of the traces along the offset lines of (8640, 10080, and 12720) m, respectively. Note that these traces pass through the salt body. Similar to Figure 5.8, all important phases in the reflectivity are well matched in the recovered image, except the phases that 128 5.8. Examples (a) (b) (c) Figure 5.6: Examples of the solution for optimization problem BPDN with SPGL1, (a) = σb, (b) = 50σb (c) = 100σb 129 5.8. Examples (a) Figure 5.7: Reflectivity model, the vertical lines are locations of investigation come from the bottom salt. Salt bottom imaging is one of the most chal- lenging fields in the seismic imaging. One reason might be the difference in the nature of the salt bottom reflector. This means the amplitude of the reflector should be interpreted as change in the acoustic impedance in that location rather than the simply perturbation in the velocity model. 5.8.2 Noisy data We examine the effectiveness of our recovery method for noisy data. For this purpose, we add white Gaussian noise to the linearized Born data with SNR = 10 dB and re-run our method on the data. Figure 5.10 shows data before and after adding the noise. Figure 5.11 shows the depth corrected migrated image for this dataset. Note that the noise is obscuring the deep reflectors in the migrated image. 130 5.8. Examples 0 500 1000 1500 2000 2500 3000 3500 4000 −10 −8 −6 −4 −2 0 2 4 6 8 Depth (m) Am pl itu de offset at 2880 (m) SPGL1 Soft Thresholding Reflectivity (a) 0 500 1000 1500 2000 2500 3000 3500 4000 −10 −8 −6 −4 −2 0 2 4 6 8 Depth (m) Am pl itu de offset at 4560 (m) SPGL1 Soft Thresholding Reflectivity (b) 0 500 1000 1500 2000 2500 3000 3500 4000 −15 −10 −5 0 5 10 Depth (m) Am pl itu de offset at 5280 (m) SPGL1 Soft Thresholding Reflectivity (c) Figure 5.8: Amplitude analysis of the recovered images using proposed meth- ods, (a) offset=3480 (m) , (b) offset=5160 (m), (c) offset=5880 (m) 131 5.8. Examples 0 500 1000 1500 2000 2500 3000 3500 4000 −12 −10 −8 −6 −4 −2 0 2 4 6 8 Depth (m) Am pl itu de offset at 8040 (m) SPGL1 Soft Thresholding Reflectivity Salt−bottom (a) 0 500 1000 1500 2000 2500 3000 3500 4000 −15 −10 −5 0 5 10 15 Depth (m) Am pl itu de offset at 9480 (m) SPGL1 Soft Thresholding Reflectivity Salt−bottom (b) 0 500 1000 1500 2000 2500 3000 3500 4000 −14 −12 −10 −8 −6 −4 −2 0 2 4 6 Depth (m) Am pl itu de offset at 12120 (m) SPGL1 Soft Thresholding Reflectivity Salt−bottom (c) Figure 5.9: Amplitude analysis of the recovered images using proposed meth- ods, (a) offset= 8640 (m), (b) offset=10080 (m) , (c) offset=12720 (m), note the salt bottom amplitude mismatch 132 5.8. Examples (a) (b) Figure 5.10: Synthetic data, (a) before , (b) after adding the noise, SNR = 10dB 133 5.8. Examples (a) Figure 5.11: Depth corrected migrated noisy image Soft-thresholding solution We investigate the effect of the threshold λ on the soft-thresholding solution for the optimization problem QP. Figure 5.12 shows the result of solving the optimization problem QP with soft thresholding for different values of the threshold λ. The solution of soft thresholding shown in Figure 5.12(a)is for when λ = σb . Figure 5.12(b) shows the result for λ = 2σb and Fig- ure 5.12(c) is for when λ = 3σb). As can be seen from the figures, by increasing λ the incoherent noise is visibly removed from the image at the expense of dimming the amplitude of the bottom reflectors. The best choise of λ is in Figure 5.12(b). BPDN solution Figure 5.13 shows the result of solving the optimization problem BPDN us- ing SPGL1 solver for different value of the tolerance . The solution of BPDN 134 5.8. Examples (a) (b) (c) Figure 5.12: Examples of the solution for optimization problem QP with soft-thresholding, (a) λ = σb, (b) λ = 2σb (c) λ = 3σb 135 5.9. Tolerance analysis is shown in Figure 5.13(a) for small ( = 10σb with σb is the standard de- viation of b in BPDN). This solution is similar to approximately inverting the approximation of normal operator in curvelet domain and applying it to migrated image (i.e., in Equation 5.11 m̃ = CTW− 1 2Cb). Figure 5.13(b) shows the result for when is larger ( = 140σb) and Figure 5.13(c) is the solution when the is high ( = 200σb). We can clearly see from the figures that when is small, the image contains high amount of noise and when the is large, the noise removed with expense of dimming the bottom reflectors. The best choice of in Figure 5.13(b) is found through exhaustive process of running the solver on a wide range of and compare the results. 5.9 Tolerance analysis In the last section, we did not specifically analyze the value of the tolerance in the optimization problem BPDN rather we empirically ran the problem with different values of and compared the results. A candidate value for the tolerance can be the statistical expectation of the misfit function in BPDN in Equation 5.8, υ̃ = E[||C− 1 2 ee (y −CTWx)||2] (5.20) with some manipulation υ̃ = N with N is the number of elements in y in lexicographical order. Indeed Figure 5.13(b) is obtained for υ = N or = Nσn. 136 5.9. Tolerance analysis (a) (b) (c) Figure 5.13: Examples of the solution for optimization problem BPDN with SPGL1, (a) = 10σb, (b) = 140σb (c) = 200σb 137 5.10. Verification of approximation 5.10 Verification of approximation Throughout our formulation, we made a subtle approximation from Equa- tion 5.8 to Equation 5.11, CTW− 1 2CCTWx ' CTW 12x. (5.21) In this approximation, we assume the projection CCT is approximately equal to identity. Here we investigate the accuracy of this approximation. We assume x = Cm with m is the true reflectivity. We call the left hand of above expression as El = CTW− 1 2CCTWx and the right hand as Er = CTW 1 2x. Figure 5.14 show the comparison between the terms in Equation 5.21. Figure 5.14(a) is the left hand El expression and Figure 5.14(b) is the right hand Er expression. The relative error between two images is 4.09 percent. 5.11 Comparison Comparing the output quality and the performance, we suggest using the soft-thresholding technique for amplitude recovery. Soft thresholding in most cases results in better or similar quality of the output with less com- putational burden comparing to BPDN. 138 5.11. Comparison (a) (b) Figure 5.14: Comparison between the approximation terms in Equation 5.21, (a) left hand term El = CTW− 1 2CCTWx , (b) right hand term Er = CTW 1 2x, relative error is 4.09 percent. 139 5.12. Conclusion 5.12 Conclusion The method presented in this chapter, is a revisit of the earlier work (Her- rmann et al. (2008)), by taking into account the migration noise and the data noise. The method relies on the curvelet transform in both approximation and inversion of the normal operator and has following features, • speeding up the calculation of the normal operator by diagonalizing it in curvelet domain. • efficiently using the compression of seismic images by curvelets to solve a curvelet-sparsity optimization problem. • bringing the amplitude correction problem within the context of stable signal recovery. The results of applying our method on synthetic data suggest that our migration amplitude recovery can be both efficient and effective in eliminat- ing migration artifacts and recovering the amplitudes in the seismic image. The recovered images showed partial elimination of noise, improved spatial resolution, and enhanced reflectivity amplitude. 5.13 Acknowledgments The authors would like to thank the authors of CurveLab for making their codes available. The authors would also like to thank W. Symes for pro- viding the migration code and TOTAL E&P for providing the synthetic BP data. This work was in part financially supported by the Natural Sciences 140 5.13. Acknowledgments and Engineering Research Council of Canada Discovery Grant (22R81254) and Collaborative Research and Development Grant DNOISE (334810-05) of Felix J. Herrmann and was carried out as part of the SINBAD project with support, secured through ITF (the Industry Technology Facilitator), from the following organizations: BG Group, BP, Chevron, ExxonMobil and Shell. 141 Bibliography Aminzadeh, F., J. Brac, and T. Kunz, 1997, 3-D Salt and Overthrust Model. SEG/EAGE 3-D Modeling Series, No. 1: Society of Exploration Geophysi- cists, Tulsa. Berg, E. V. D. and M. Friedlander, 2007, In pursuit of a root: Optimization Online:2007.06.1708, 8. Beylkin, G., 1985, Imaging of the discontinuities in the inverse scattering problem by inversion of casual generalized curvelet transform: Journal of Mathematical Physics, 26, 99–108. Candès, E. J., J. Romberg, and T. Tao, 2006, Stable signal recovery from incomplete and inaccurate measurements: 59, 1207–1223. Chattopadhyay, S. and G. A. McMechan, 2008, Imaging conditions for prestack reverse-time migration: Geophysics, 73, S81–S89. Claerbout, J., 1985, Earth soundings analysis, processing versus inversion: Blackwell Scientific Publications. Daubechies, I., M. Defrise, and C. de Mol, 2005, An iterative thresholding algorithm for linear inverse problems with a sparsity constraints: 57, 1413–1457. Donoho, D., 1995, Nonlinear solution of linear inverse problems by wavelet- vaguelette decomposition: Applied and Computational Harmonic Analy- 142 Chapter 5. Bibliography sis 2, 101 – 126. E. Candes, L. Demanet, D. D. L. Y., 2005, Fast discrete curvelet transforms. E. J. Cand‘es, L. Demanet, D. L. D. and L. Ying, 2006, Fast discrete curvelet transforms: SIAM Multiscale Model. Simul., 5, 861899. Gray, S., 1997, True amplitude seismic migration: A comparison of three approaches: Geophysics, 62, 929–936. Guitton, A., 2004, Amplitude and kinematic corrections of migrated images for nonunitary imaging operators: Geophysics, 69, 1017–1024. Hennenfent, G. and F. J. Herrmann, 2006, Seismic denoising with non- uniformly sampled curvelets: Comp. in Sci. and Eng., 8, 16–25. Herrmann, F., C. Brown, Y. Erlangga, and P. Moghaddam, 2009, Curvelet- based migration preconditioning and scaling: Geophysics, 74, A41–A45. Herrmann, F., P. Moghaddam, and C. Stolk, 2008, Sparsity- and continuity- promoting seismic image recovery with curvelet frames: Applied and Com- putational Harmonic Analysis 24, 2, 150–173. Herrmann, F. J., D. Wang, and D. J. Verschuur, 2007, Adaptive curvelet- domain primary-multiple separation. ((Submitted to Geophysics)). J. C. Costa, F. A. Silva Neto, M. R. M. A. J. S. and A. Novais, 2009, Obliquity-correction imaging condition for reverse time migration: Geo- physics, 74, S57–S66. J. Schleicher, J. C. C. and A. Novais, 2008, A comparison of imaging condi- tions for wave-equation shot-prole migration: Geophysics, 73, S219–S227. Kuehl, H. and M. Sacchi, 2003, Least-squares wave-equation migration for AVP/AVA inversion: Geophysics, 68, 262–273. Mulder, W. and R.-E. Plessix, 2004, A comparison between one-way and 143 Chapter 5. Bibliography two-way wave equation migration: Geophysics, 69, 14911504. Nemeth, T., C. Wu, and G. Schuster, 1999, Least-squares migration of in- complete reflection data: Geophysic, 64, 208–221. Nocedal, J. and W. Wright, 1999, Numerical optimization: Springer Verlag. O’Brien, M. and S. Gray, 1996, Can we image beneath salt?: The Leading Edge, 15, 17–22. Plessix, R. and W. Mulder, 2004, Frequency-domain finite-difference amplitude-preserving migration: Geophys. J. Int., 157, 975–987. Rakesh, A., 1988, A linearized inverse problem for the wave equation: Comm. on P.D.E., 13, 573–601. Rickett, J., 2003, Illumination based normalization for wave-equation depth migration: Geophysics, 68, 208–221. Shahidi, R. and F. Herrmann, 2009a, Curvelet-domain matched filtering: under preparation. ——–, 2009b, Curvelet-domain matched filtering with frequency-domain regularization: 28. Stolk, C., 2000, Microlocal analysis of a seismic linearized inverse problem: Wave Motion, 32, 267–290. Stolk, C. and M. De-Hoop, 2006, Seismic inverse scattering in the downward continuation approach: Wave Motion, 43, 579–598. Symes, W. W., 2007, Reverse time migration with optimal checkpointing: Geophysics, 72, SM213–SM221. ——–, 2008a, Approximate linearized inversion by optimal scaling of prestack depth migration: Geophysics, 73, R23–R35. ——–, 2008b, Migration velocity analysis and waveform inversion: Geophys- 144 Chapter 5. Bibliography ical Prospecting, 56, 765 – 790. Tarantola, A., 1987, Inverse problem theory: Elseweir. Ten-Kroode, A., D. Smith, and A. Verdel, 1998, A microlocal analysis of migration: Wave Motion, 28. van den Berg, E. and M. P. Friedlander, 2008, Probing the pareto frontier for basis pursuit solutions: SIAM Journal on Scientific Computing, 31, 890–912. William W. Symes, Christiaan C. Stolk, B. B. and F. Gao, 2004, Reverse time shot-geophone migration: Technical Report. Y. Zhang, S. Xu, N. B. and G. Zhang, 2007, True-amplitude, angle-domain, common-image gathers from one-way wave-equation migrations: Geo- physics, 72, S49–S58. 145 Chapter 6 Conclusions In this chapter I summarize the main contributions of this thesis and discuss some limitations of the work presented. I also suggest follow-up work as well as possible extensions. 6.1 Main contributions The topic of this thesis is seismic image amplitude correction which is directly applied to seismic imaging. The approach I advocate is to view seis- mic imaging in a mathematical perspective. I identify a transform, called the curvelet transform (Candès and Donoho, 2004), to that effect and use it in a new formulation of the seismic images amplitude correction problem. The most challenging operator in the seismic imaging problem, called nor- mal operator, is a computationally exhausting. This operator needs to be inverted during the imaging process to give an accurate understanding about the reflectors amplitude in a seismic image. Each element of the curvelet transform, namely curvelet, has an attractive property under the normal operator. In this thesis, I proved that each curvelet remains approximately invariant under the normal operator, which gives us an exceptional ability to approximately invert the normal operator with only one iteration. I call 146 6.1. Main contributions this approach the curvelet match filtering (CMF). I leverage this approximation towards the development of a stable am- plitude recovery by coining it with sparsity-promoting inversion, which is basically a large-scale one-norm solver. The remainder of this section provides more details about the contribu- tions. 6.1.1 Sparsity and continuity promoting seismic image recovery with curvelet frames I combine the compression of images by curvelets with the invariance of this transform under the normal operator. This combination allows for a formulation of a stable recovery method for seismic amplitudes. During the recovery, the normal operator is approximately inverted. Compared to other approaches for migration scaling, the presented method (i) includes a theoretical bound on the L2 − error for the diagonal approximation in the curvelet domain; (ii) prescribes a procedure for the estimation of the diag- onal from numerical implementations of the imaging operators; (iii) formu- lates the amplitude recovery problem as a nonlinear optimization problem, where the inversion of the diagonalized normal operator is regularized by imposing sparsity in the curvelet domain and continuity along the imaged reflectors. The diagonal approximation is used to formulate the seismic amplitude recovery in terms of a constrained optimization problem. The amplitude- corrected image is obtained by solving this sparsity- and continuity-promoting optimization problem. The invariance of curvelets under the normal opera- 147 6.1. Main contributions tor preserves the sparsity. The cost of computing the diagonal approxima- tion is one demigration-migration per reference vector, which is much less compared to the cost of Krylov-based least-squares inversion. The recov- ery results show an overall improvement of the image quality. The joined sparsity- and continuity-enhanced image has diminished artifacts, improved resolution and recovered amplitudes. 6.1.2 Curvelet-based migration preconditioning and scaling I present a method to approximate the normal operator and use this ap- proximation as a preconditioner in the least squares framework. Because of the size of the seismic imaging problem, preconditioning of least-squares migration is an interesting topic. Traditionally, the first few iterations of the LSQR algorithm of least-squares migration are known to make significant progress towards the solution. Unfortunately, even for this limited number of iterations, the computational costs are often still prohibitively large for practical problems. Our method, partly resolves this issue through a com- bination of left and right preconditioning together with a curvelet-domain scaling. Inclusion of the latter proved particularly important because it re- stores the amplitudes and leads to faster convergence, at a relatively small computational overhead. Preconditioning also plays a pivotal role in mak- ing this approach numerically feasible and may extend to a solution of the full-waveform inversion problem with sparsity promotion. Both approaches are justified by ample evidence that curvelets are sparse on the model. 148 6.1. Main contributions 6.1.3 Curvelet-based seismic data processing Beside the seismic wavefield reconstruction problem, I recast a few other processing steps—signal separation, migration-amplitude recovery, and deconvolution— in a sparsity-promoting program that exploits the high degree of sparsity attained by curvelets on seismic data and images. The promising results obtained show that the insights gained from the developments of CMF can be leveraged towards a much broader range of applications. This prospect opens an exciting new outlook towards future developments in exploration seismology. 6.1.4 True amplitude depth migration using curvelets I presented a fast and robust approach for approximation of the normal operator by revisiting of the earlier work and taking into account the migra- tion noise and the data noise. Compared to other approaches for migration amplitude recovery, some improvements in the design of curvelet amplitude recovery method including, speeding up the calculation of the normal oper- ator by diagonalizing it in curvelet domain, designing an efficient approxi- mation that takes into account a laterally-variant velocity models and steep reflectors. The results of applying method on synthetic data suggest that migration amplitude recovery method can be both efficient and effective in eliminating migration artifacts and recovering the amplitudes in the seismic image. The recovered images showed partial elimination of noise, improved spatial resolution, and enhanced reflectivity amplitude. 149 6.2. Follow-up work 6.2 Follow-up work I suggest a few ideas that go beyond the reported experiments. 6.2.1 Full waveform inversion Our estimation for normal operator can be directly applied to the context of full waveform inversion in which the normal operator is inverted and used iteratively. In this context, the velocity model (not just reflectivity in the migration context) is estimated directly from data (see e.g. Symes, 2008). 6.2.2 3D true amplitude migration Our 2D true amplitude migration can be easily extend to 3D. This can be done in two ways, first, I apply the curvelet smoothing in 2D and for 3rd dimension I can exploit wavelet smoothing. Second, I can use the 3D curvelet transform and extend the method to 3D curvelet diagonal estimation. The latter, needs a sophisticated algorithm to locate the closest curvelets in 3D rather than 2D for smoothing operator. 6.3 Adding more priori knowledge to solution In this thesis, I propose some ideas for amplitude recovery which pre- sume that the seismic image are sparse in the curvelet domain. The common theme of most ideas is that there is no priori knowledge of the geological structure of the survey area. However, if a priori knowledge about the struc- ture already obtained through the well-log or other geological/geophysical 150 6.4. Current limitations experiments, it can be used within stable recovery framework in addition of sparsity to reconstruct seismic wavefields. 6.4 Current limitations I examine both the practical and the fundamental weaknesses of the current curvelet transform. 6.4.1 Curvelet code The true amplitude migration results presented in this thesis were ob- tained using the FDCT based on the wrapping of specially selected Fourier samples (Candès et al., 2005). This implementation breaks down the input image or volume into a number of scales depending on the length of the shortest axis. In other words, if one axis is much shorter than the others, the decomposition along the long axes is unnecessarily limited. Despite an increased implementation complexity, an alternative would be to extend the short axis to be sufficiently large enough. This extension can be done by padding the image along the short axis or wrapping it. A more fundamental limitation of the FDCT is related to the redun- dancy of the transform. Indeed, the FDCT is around 8-redundant in 2D and around 24-redundant in 3D, which precludes, at least for now, tractable higher-dimensional FDCTs. Lu and Do (2007) propose a less redundant N - dimensional (N ≥ 2) implementation, termed surfacelet transform, by com- bining a directional filter bank with a multiscale pyramid. Another option is to combine the curvelet transform with another transform (Herrmann, 2003; 151 6.4. Current limitations Neelamani et al., 2008) to reduce redundancy and reach higher dimensions. The different treatment of the axes is unsatisfactory in several applications (see, e.g., Neelamani et al., 2008), though. For interest, Kutyniok and La- bate (2005) propose yet another N -dimensional (N ≥ 2) transform, called shearlet transform, but no discrete implementation is available at this point to determine the redundancy and the effectiveness of shearlets for wavefield reconstruction. In addition, in some cases, defining the smoothing opera- tor in the transform domain is not as straightforward as curvelets since in some cases finding the closest elements in transform domain might require a significant amount of operations. 152 Bibliography Candès, E. J., L. Demanet, D. L. Donoho, and L. Ying, 2005, Curvelab: software. (Available at http://www.curvelet.org). Candès, E. J. and D. L. Donoho, 2004, New tight frames of curvelets and optimal representations of objects with C2 singularities: Communications on Pure and Applied Mathematics, 57, no. 2, 219 – 266. Herrmann, F. J., 2003, Multifractional splines: Application to seismic imag- ing. Kutyniok, G. and D. Labate, 2005, Shearlets: website. (www.shearlet.org). Lu, Y. M. and M. N. Do, 2007, Multidimensional directional filter banks and surfacelets: Transactions on Image Processing, 16, no. 4. Neelamani, R., A. I. Baumstein, D. G. Gillard, M. T. Hadid, and W. L. Soroka, 2008, Coherent and random noise attenuation using curvelet transforms: The Leading Edge. (In press). Symes, W. W., 2008, Migration velocity analysis and waveform inversion: Geophysical Prospecting, 56, no. 6, 765–790. 153 Appendix A Proofs of lemma 1 and theorem 1 We first prove Theorem 1 from Lemma 1, we then prove the Lemma which appears in chapter 2. Proof: [Proof of Theorem 1] CT acting on a vector c = {cµ}µ∈M in `2 is simply given by CTc = ∑ ν cνϕν . Hence, CTDΨCϕµ = ∑ ν (Cϕµ)νa(xν , ξν)ϕν . The difference (Ψ(x,D)−CTDΨC)ϕµ can be written as, using that CTC = Id, ∑ ν (Cϕµ)ν(Ψ(x,D)− a(xν , ξν))ϕν The vector (Cϕµ)ν is bounded in `2(M) of vectors with curvelet indices, and its entries can only be nonzero if |µ| − 2 ≤ |ν| ≤ |µ| + 2, since the support of two curvelets in the Fourier domain is disjoint when curvelets are two or A version of this appendix has been published. Herrman, F.J., Moghaddam, P.P. and Stolk, C. (2008) Sparsity- and continuity-promoting seismic image recovery with curvelet frames. Applied and Computational Harmonic Analysis, Vol. 24, No. 2, pp. 150-173, 2008. 154 Appendix A. Proofs of lemma 1 and theorem 1 more scales apart. It follows that ‖(Ψ(x,D)− CTDΨC)ϕµ‖2L2(Rn) ≤ C6 ∑ ν ‖(Cϕµ)ν(A(x,D)− a(xν , ξν))ϕν‖2 L2(Rd) ≤ C7 ∑ ν 2−|ν||(Cϕµ)ν |2 ≤ C82−|µ|/2. Proof: [Proof of Lemma 1] Possibly after a coordinate transformation, which does not affect the pseudodifferential nature of Ψ, we may assume that ξµ = (0, . . . , 0, λ), we will write ξ′ = (ξ1, . . . , ξd−1). We may take |ν| greater than some minimum value, so that the support in the wavenumber domain is away from ξ = 0. We can then assume that the principal symbol is homogeneous of order 0 in the support of the curvelet. On the support of the curvelet in the Fourier domain the symbol can be written as a(x, ξ ′ ξn ). We apply a preparation theorem to this term. There are C∞ functions bn(x, ξ′/ξd), n = 1, . . . , 2d− 1, such that a(x, ξ′/ξd)− a(xν , ξ′ν/ξν,d) = d∑ n=1 (x− xν)nbn(x, ξ′/ξd) + d−1∑ n=1 ( ξn ξd − ξν,n ξν,d ) bd+n(x, ξ′/ξd). Therefore, we can write ( Ψ(x,D′/Dd)− a(xν , ξ′ν/ξν,d) ) ϕν = [ d∑ n=1 (x− xν)nbn(x,D′/Dd) + d−1∑ n=1 bd+n(x,D′/Dd) Dn Dd ] ϕν . (A.1) Consider first the contribution for one of the bn with 1 ≤ n ≤ d. The 155 Appendix A. Proofs of lemma 1 and theorem 1 operator acting on the curvelet reads (x− xν)nbn(x,D′/Dd)ϕν = bn(x,D′/Dd)(x− xν)nϕν + [(x− xν)n, bn(x,D′/Dd)]ϕν . Because of the support and decay properties of curvelets we have ‖(x − xν)nϕν‖L2 ≤ C−|ν|/29 , therefore ‖bn(x,D′/Dd)(x− xν)nϕν‖L2(Rd) ≤ C102 −|ν|/2 By the calculus of pseudodifferential operators, the operator [(x−xν)n, bn(x,D′/Dd)] is a pseudodifferential operator of order −1, which is continuous H−1(Rd) to L2(Rd) therefore ‖[(x− xν)n, bn(x,D′/Dd)]ϕν‖L2(Rd) ≤ C10‖ϕν‖H−1(Rd) ≤ C112 −|ν|. Adding the previous two estimates, we find that ‖(x− xν)nbn(x,D′/Dd)ϕν‖L2(Rd) ≤ C122 −|ν|/2. (A.2) We now consider the bn, d+1 ≤ n ≤ 2d−1. We must estimate bd+n(x,D′/Dd)DnDdϕν . By (2.11) we have ‖DnDdϕν‖L2(Rn) ≤ C132−|ν|/2. Therefore, ‖bd+n(x,D′/Dd)D ′ Dd ϕν‖L2(Rd) ≤ C142 −|ν|/2. (A.3) From (A.1) and the estimates (A.2) and (A.3) for the summands, we find 156 Appendix A. Proofs of lemma 1 and theorem 1 the result. 157 Appendix B Reverse time wave equation migration This appendix is part of chapter 5. Most migration operators are defined to be the adjoint of what is called the scattering operator. This assumption is also true for reverse time migration. B.1 Single scattering The causal acoustic Green’s function G(x, t;xs),x ∈ R3 for a point source at x = xs is the solution of 1 v2(x) ∂2G ∂t2 (x, t;xs)−∇2xG(x, t;xs) = δ(x− xs)δ(t) (B.1) with G = 0, t < 0 and v is the acoustic wave velocity field. Denote by m(x) = δv(x)/v(x) a relative perturbation of the velocity field. Then linearization of the wave equation yields for the corresponding perturbation of the Green’s function A version of this appendix will be submitted for publication. Moghaddam, P.P., Herrmann F. J. and Shahidi, R. (2010) True amplitude depth migration using curvelets. 158 B.2. Shot-geophone modeling and migration 1 v2(x) ∂2δG ∂t2 (x, t;xs)−∇2xδG(x, t;xs) = 2m(x) v2(x) ∂2 ∂t2 G(x, t;xs) (B.2) whose solution has the integral representation at the source and receiver points xr,xs δG(xr, t;xs) = ∂2 ∂t2 ∫ dx ∫ dτ 2m(x) v2(x) G(x, t− τ ;xr)G(x, τ ;xs) (B.3) B.2 Shot-geophone modeling and migration The single-scattered wave field is the time convolution of δG with a source wavelet. The main concern of this paper is the kinematic relationships between data and image, thus we ignore the filtering effect of the source functional and replace it by delta function. This replacement of the source by an impulse does not violate any of our assumptions regarding the adjoint state method, thus the Born modeling operator K[v] is K[v]m(x) = δG(xr, t;xs) (B.4) The crux of our amplitude recovery method relies on the shot-geophone migration operator to be the adjoint of the shot-geophone modeling opera- tor. The derivation of the adjoint reverse time implementation is a minor variation on the usual implementation of reverse time migration (the ’ad- joint state method’, ((see e.g. Tarantola, 1987; Whitmore, 1983; Lailly, 1983; Yoon and Hong, 2003; Symes and Stolk, 2004; Symes, 2007)). The 159 B.2. Shot-geophone modeling and migration result is K∗[v]d(xr, t;xs) = m̂(x) = − ∫ dxs ∫ T 0 dt2v(x)q(x, t;xs) ∂2G ∂t2 (x, t;xs) (B.5) where the adjoint state or backpropagated field q(x, t;xs) satisfies q = 0, t > T and 1 v2(x) ∂2q ∂t2 (x, t;xs)−∇2xq(x, t;xs) = ∫ dxrd(xr, t;xs)δ(x− xr) (B.6) The migration operator defined by above equations is the reverse time migration operator. Symes and Stolk (2004) showed that the migration operator which is defined in equations B.5 and B.6 is the adjoint of the modeling operator defined in equation B.3. The reverse time migration that is used for this work is based on the Symes and Stolk (2004) implementation, for which the adjointness of the modeling and migration operators is properly tested in the discrete sense. By having the migration and modeling operators properly set, we can proceed with our amplitude recovery method. 160 Bibliography Lailly, P., 1983, The seismic inverse problem as a sequence of before stack migrations: SIAM, 206–220. Symes, W., 2007, Reverse time migration with optimal checkpointing: Geo- physics, 72, 213–221. Symes, W. and C. Stolk, 2004, Reverse time shot-geophone migration: The Rice Inversion Project, Department of Computational and Applied Math- ematics, Rice University, Houston, Texas, USA. Tarantola, A., 1987, Inverse problem theory: Elseweir. Whitmore, N. D., 1983, Iterative depth migration by backward time propa- gation: Presented at 53rd Annual International SEG Meeting, Las Vegas. Yoon, K. and S. Hong, 2003, 3D reverse-time migration using hte acous- tic wave equation: An experience with the SEG/EAGE data set.: The Leading Edge, 22–38. 161
Cite
Citation Scheme:
Usage Statistics
Country | Views | Downloads |
---|---|---|
Japan | 7 | 0 |
Russia | 3 | 0 |
India | 3 | 0 |
United States | 1 | 0 |
City | Views | Downloads |
---|---|---|
Tokyo | 7 | 0 |
Unknown | 3 | 0 |
Pune | 3 | 0 |
Ashburn | 1 | 0 |
{[{ mDataHeader[type] }]} | {[{ month[type] }]} | {[{ tData[type] }]} |
Share
Share to: