imaging: adaptive ground roll removal · 2019-11-14 · number of di! erent ori entations. using...

37
Curvelet Processing and Imaging: Adaptive Ground Roll Removal The University of British Columbia Department of Earth and Ocean Sciences *Veritas DGC Inc. Carson Yarham, Felix Herrmann, and Daniel Trad* Released to public domain under Creative Commons license type BY (https://creativecommons.org/licenses/by/4.0). Copyright (c) 2004 SLIM group @ The University of British Columbia.

Upload: others

Post on 19-Mar-2020

1 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

Curvelet Processing and Imaging: Adaptive

Ground Roll Removal

The University of British ColumbiaDepartment of Earth and Ocean Sciences

*Veritas DGC Inc.

Carson Yarham, Felix Herrmann, and Daniel Trad*

Released to public domain under Creative Commons license type BY (https://creativecommons.org/licenses/by/4.0).Copyright (c) 2004 SLIM group @ The University of British Columbia.

Page 2: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

Oz25 Data Set from Yilmaz’s Seismic Data Processing

The Problem

200

400

600

800

1000

20 40 60 80

Page 3: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

What We Can Do

200

400

600

800

1000

20 40 60 80

200

400

600

800

1000

20 40 60 80

Page 4: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

• Represented as a Rayleigh wave.

• Dispersive, Low Frequency

• Highly dependent on near surface properties

• Reduces SNR

• Generated in the frequency slope domain in the slant stack transform

Ground Roll Properties

100

200

300

400

500

50 100

(A.G. McMechan and M.J. Yeldin, Geophysics, 1981)

Page 5: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

Two Problems to Solve

How Do We Identify What to Remove?

How Do We Remove It?

- Modeled Ground Roll- Noise Prediction From Other Methods

- Incorporate Prior Predictions- Use Adaptive Subtraction

Page 6: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

• Identifies hyperbolic reflectors from the signal

• May produce artifacts with conventional subtraction

• We can use the predicted noise with adaptive subtraction

Our Use Of Hyperbolic Radon Filtering

AdaptiveSubtraction

PredictedNoise

NoisyData Result

Page 7: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

• Wavelets:

• Represent time and frequency

• Multi-Scale

• Curvelets:

• Local in position and angle

• Strongly anisotropic at fine scales (parabolic scaling principle: length² ~ width)

Wavelets and Curvelets

Candes 02, Do 02Candes 02, Do 02Candes 02, Do 02

Seismic texture analysis by curve modelling with applications to facies analysisErik Monsen∗ ([email protected]), Stavanger University College, Stavanger, NorwayJan E. Odegard ([email protected]), Rice University, Houston, Texas, USA

Summary

Seismic facies analysis and seismic texture analysis canbe seen as nearly equivalent operations, implying thatif we are able to solve the texture analysis problem,corresponding seismic facies labels may subsequentlybe assigned. In this paper we consider the use ofa directional multiresolution transform, the curvelettransform, for seismic texture analysis. This transformis optimal for curves in 2D as it obeys the inherent curvescaling relation. Preliminary results on seismic textureclassification are very promising.

Introduction

Seismic facies analysis (7) is based on considerationsof three primary ingredients, internal texture (parallel,subparallel, chaotic, etc.), external shape (sheet, drape,mound, etc) and boundary conditions (onlap, toplap,downlap, etc.). The two latter parts, boundary andshape, are to a large extent given as a function of theinternal texture’s extent. Hence, the focus in this paperwill be on texture analysis.

Fundamental to the formation of the structures seen inseismic images is the process of deposition. On its own,deposition creates uniform, parallel layers of rock. How-ever, the dynamics of the earth’s crust deforms this basicstructure in a variety of ways producing other compositestructures. If we were to break seismic images down totheir smallest elements of parallel structures, or line/curvesegments, these small elements could be used as a ba-sis for seismic facies analysis/classification. Towards thisend, we will consider a new member of the family of mul-tiresolution transforms, the curvelet transform.

Multiresolution transforms have proven to be very suc-cessful in texture processing applications, be it for anal-ysis or synthesis purposes (8). Their success can be at-tributed to the fact that textures often exhibit details dis-tributed over a wide range of scales, and that the trans-forms often have manageable computational costs. In thisfamily the wavelet transform has received a lot of atten-tion due to its ability to sparsely represent signals andimages producing impressive results in the area of com-pression (9) and denoising (6).

The wavelet transform can in general be seen as a mul-tiscale edge detector, be it for 1-D, 2-D or higher di-mensional signals. For 1-D piecewise smooth signals thetransform is easily seen to be sparse. Consequently, non-linear approximation of such signals using wavelets arevery powerful.

However, for 2-D piecewise smooth signals a sparse rep-

(a)

(b)

Fig. 1: Using the wavelet and curvelet bases for representingedges in 2-D images results in very different approximationperformances (5). (a) Approximation using a wavelet basis,and (b) a curvelet basis.

resentations is not implied. The primary reason for thislack of sparsity stems from the fact that a separable imple-mentation of the 2-D wavelet transform is typically beingused. Using 1-D transforms in the two directions, whereeach transform sparsely approximates edges, results in a2-D transform that is good at representing isolated dis-continuities in the images. However, since discontinu-ities in natural images, and seismic images, tend to occuralong lines/contours, thus, exhibiting high local geometri-cal correlation, the separable wavelet transform becomes aredundant representation. That is, the wavelet transformis no longer optimal since we now could use basis func-tions having wider support along the direction of highestgeometrical correlation. This is best explained visually(5) if we try to approximate a curved discontinuity in 2-D space by “painting” with wavelets having square sup-port. Using dyadic scaling, the best set of brush strokesis seen in Figure 1(a). Suppose now that we have an-other basis/“brush” that has rectangular support along anumber of different orientations. Using the second basis,the curvelet basis (5), we are able to better approximatethe curve, as seen in Figure 1(b). This transform is truly2-D, not a construction from 1-D transforms. The result-ing basis functions can detect small segments of the curve

Candes 00, Donoho 95, Do 01

Page 8: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

Candes ‘02

Curvelets

Seismic texture analysis by curve modelling with applications to facies analysisErik Monsen∗ ([email protected]), Stavanger University College, Stavanger, NorwayJan E. Odegard ([email protected]), Rice University, Houston, Texas, USA

Summary

Seismic facies analysis and seismic texture analysis canbe seen as nearly equivalent operations, implying thatif we are able to solve the texture analysis problem,corresponding seismic facies labels may subsequentlybe assigned. In this paper we consider the use ofa directional multiresolution transform, the curvelettransform, for seismic texture analysis. This transformis optimal for curves in 2D as it obeys the inherent curvescaling relation. Preliminary results on seismic textureclassification are very promising.

Introduction

Seismic facies analysis (7) is based on considerationsof three primary ingredients, internal texture (parallel,subparallel, chaotic, etc.), external shape (sheet, drape,mound, etc) and boundary conditions (onlap, toplap,downlap, etc.). The two latter parts, boundary andshape, are to a large extent given as a function of theinternal texture’s extent. Hence, the focus in this paperwill be on texture analysis.

Fundamental to the formation of the structures seen inseismic images is the process of deposition. On its own,deposition creates uniform, parallel layers of rock. How-ever, the dynamics of the earth’s crust deforms this basicstructure in a variety of ways producing other compositestructures. If we were to break seismic images down totheir smallest elements of parallel structures, or line/curvesegments, these small elements could be used as a ba-sis for seismic facies analysis/classification. Towards thisend, we will consider a new member of the family of mul-tiresolution transforms, the curvelet transform.

Multiresolution transforms have proven to be very suc-cessful in texture processing applications, be it for anal-ysis or synthesis purposes (8). Their success can be at-tributed to the fact that textures often exhibit details dis-tributed over a wide range of scales, and that the trans-forms often have manageable computational costs. In thisfamily the wavelet transform has received a lot of atten-tion due to its ability to sparsely represent signals andimages producing impressive results in the area of com-pression (9) and denoising (6).

The wavelet transform can in general be seen as a mul-tiscale edge detector, be it for 1-D, 2-D or higher di-mensional signals. For 1-D piecewise smooth signals thetransform is easily seen to be sparse. Consequently, non-linear approximation of such signals using wavelets arevery powerful.

However, for 2-D piecewise smooth signals a sparse rep-

(a)

(b)

Fig. 1: Using the wavelet and curvelet bases for representingedges in 2-D images results in very different approximationperformances (5). (a) Approximation using a wavelet basis,and (b) a curvelet basis.

resentations is not implied. The primary reason for thislack of sparsity stems from the fact that a separable imple-mentation of the 2-D wavelet transform is typically beingused. Using 1-D transforms in the two directions, whereeach transform sparsely approximates edges, results in a2-D transform that is good at representing isolated dis-continuities in the images. However, since discontinu-ities in natural images, and seismic images, tend to occuralong lines/contours, thus, exhibiting high local geometri-cal correlation, the separable wavelet transform becomes aredundant representation. That is, the wavelet transformis no longer optimal since we now could use basis func-tions having wider support along the direction of highestgeometrical correlation. This is best explained visually(5) if we try to approximate a curved discontinuity in 2-D space by “painting” with wavelets having square sup-port. Using dyadic scaling, the best set of brush strokesis seen in Figure 1(a). Suppose now that we have an-other basis/“brush” that has rectangular support along anumber of different orientations. Using the second basis,the curvelet basis (5), we are able to better approximatethe curve, as seen in Figure 1(b). This transform is truly2-D, not a construction from 1-D transforms. The result-ing basis functions can detect small segments of the curve

Seismic texture analysis by curve modelling with applications to facies analysisErik Monsen∗ ([email protected]), Stavanger University College, Stavanger, NorwayJan E. Odegard ([email protected]), Rice University, Houston, Texas, USA

Summary

Seismic facies analysis and seismic texture analysis canbe seen as nearly equivalent operations, implying thatif we are able to solve the texture analysis problem,corresponding seismic facies labels may subsequentlybe assigned. In this paper we consider the use ofa directional multiresolution transform, the curvelettransform, for seismic texture analysis. This transformis optimal for curves in 2D as it obeys the inherent curvescaling relation. Preliminary results on seismic textureclassification are very promising.

Introduction

Seismic facies analysis (7) is based on considerationsof three primary ingredients, internal texture (parallel,subparallel, chaotic, etc.), external shape (sheet, drape,mound, etc) and boundary conditions (onlap, toplap,downlap, etc.). The two latter parts, boundary andshape, are to a large extent given as a function of theinternal texture’s extent. Hence, the focus in this paperwill be on texture analysis.

Fundamental to the formation of the structures seen inseismic images is the process of deposition. On its own,deposition creates uniform, parallel layers of rock. How-ever, the dynamics of the earth’s crust deforms this basicstructure in a variety of ways producing other compositestructures. If we were to break seismic images down totheir smallest elements of parallel structures, or line/curvesegments, these small elements could be used as a ba-sis for seismic facies analysis/classification. Towards thisend, we will consider a new member of the family of mul-tiresolution transforms, the curvelet transform.

Multiresolution transforms have proven to be very suc-cessful in texture processing applications, be it for anal-ysis or synthesis purposes (8). Their success can be at-tributed to the fact that textures often exhibit details dis-tributed over a wide range of scales, and that the trans-forms often have manageable computational costs. In thisfamily the wavelet transform has received a lot of atten-tion due to its ability to sparsely represent signals andimages producing impressive results in the area of com-pression (9) and denoising (6).

The wavelet transform can in general be seen as a mul-tiscale edge detector, be it for 1-D, 2-D or higher di-mensional signals. For 1-D piecewise smooth signals thetransform is easily seen to be sparse. Consequently, non-linear approximation of such signals using wavelets arevery powerful.

However, for 2-D piecewise smooth signals a sparse rep-

(a)

(b)

Fig. 1: Using the wavelet and curvelet bases for representingedges in 2-D images results in very different approximationperformances (5). (a) Approximation using a wavelet basis,and (b) a curvelet basis.

resentations is not implied. The primary reason for thislack of sparsity stems from the fact that a separable imple-mentation of the 2-D wavelet transform is typically beingused. Using 1-D transforms in the two directions, whereeach transform sparsely approximates edges, results in a2-D transform that is good at representing isolated dis-continuities in the images. However, since discontinu-ities in natural images, and seismic images, tend to occuralong lines/contours, thus, exhibiting high local geometri-cal correlation, the separable wavelet transform becomes aredundant representation. That is, the wavelet transformis no longer optimal since we now could use basis func-tions having wider support along the direction of highestgeometrical correlation. This is best explained visually(5) if we try to approximate a curved discontinuity in 2-D space by “painting” with wavelets having square sup-port. Using dyadic scaling, the best set of brush strokesis seen in Figure 1(a). Suppose now that we have an-other basis/“brush” that has rectangular support along anumber of different orientations. Using the second basis,the curvelet basis (5), we are able to better approximatethe curve, as seen in Figure 1(b). This transform is truly2-D, not a construction from 1-D transforms. The result-ing basis functions can detect small segments of the curve

Seismic texture analysis by curve modelling with applications to facies analysisErik Monsen∗ ([email protected]), Stavanger University College, Stavanger, NorwayJan E. Odegard ([email protected]), Rice University, Houston, Texas, USA

Summary

Seismic facies analysis and seismic texture analysis canbe seen as nearly equivalent operations, implying thatif we are able to solve the texture analysis problem,corresponding seismic facies labels may subsequentlybe assigned. In this paper we consider the use ofa directional multiresolution transform, the curvelettransform, for seismic texture analysis. This transformis optimal for curves in 2D as it obeys the inherent curvescaling relation. Preliminary results on seismic textureclassification are very promising.

Introduction

Seismic facies analysis (7) is based on considerationsof three primary ingredients, internal texture (parallel,subparallel, chaotic, etc.), external shape (sheet, drape,mound, etc) and boundary conditions (onlap, toplap,downlap, etc.). The two latter parts, boundary andshape, are to a large extent given as a function of theinternal texture’s extent. Hence, the focus in this paperwill be on texture analysis.

Fundamental to the formation of the structures seen inseismic images is the process of deposition. On its own,deposition creates uniform, parallel layers of rock. How-ever, the dynamics of the earth’s crust deforms this basicstructure in a variety of ways producing other compositestructures. If we were to break seismic images down totheir smallest elements of parallel structures, or line/curvesegments, these small elements could be used as a ba-sis for seismic facies analysis/classification. Towards thisend, we will consider a new member of the family of mul-tiresolution transforms, the curvelet transform.

Multiresolution transforms have proven to be very suc-cessful in texture processing applications, be it for anal-ysis or synthesis purposes (8). Their success can be at-tributed to the fact that textures often exhibit details dis-tributed over a wide range of scales, and that the trans-forms often have manageable computational costs. In thisfamily the wavelet transform has received a lot of atten-tion due to its ability to sparsely represent signals andimages producing impressive results in the area of com-pression (9) and denoising (6).

The wavelet transform can in general be seen as a mul-tiscale edge detector, be it for 1-D, 2-D or higher di-mensional signals. For 1-D piecewise smooth signals thetransform is easily seen to be sparse. Consequently, non-linear approximation of such signals using wavelets arevery powerful.

However, for 2-D piecewise smooth signals a sparse rep-

(a)

(b)

Fig. 1: Using the wavelet and curvelet bases for representingedges in 2-D images results in very different approximationperformances (5). (a) Approximation using a wavelet basis,and (b) a curvelet basis.

resentations is not implied. The primary reason for thislack of sparsity stems from the fact that a separable imple-mentation of the 2-D wavelet transform is typically beingused. Using 1-D transforms in the two directions, whereeach transform sparsely approximates edges, results in a2-D transform that is good at representing isolated dis-continuities in the images. However, since discontinu-ities in natural images, and seismic images, tend to occuralong lines/contours, thus, exhibiting high local geometri-cal correlation, the separable wavelet transform becomes aredundant representation. That is, the wavelet transformis no longer optimal since we now could use basis func-tions having wider support along the direction of highestgeometrical correlation. This is best explained visually(5) if we try to approximate a curved discontinuity in 2-D space by “painting” with wavelets having square sup-port. Using dyadic scaling, the best set of brush strokesis seen in Figure 1(a). Suppose now that we have an-other basis/“brush” that has rectangular support along anumber of different orientations. Using the second basis,the curvelet basis (5), we are able to better approximatethe curve, as seen in Figure 1(b). This transform is truly2-D, not a construction from 1-D transforms. The result-ing basis functions can detect small segments of the curve

Candes 02, Do 02

1403 NOTICES OF THE AMS VOLUME 50, NUMBER 11

localized in just a few coefficients. This can bequantified. Simply put, there is no basis in whichcoefficients of an object with an arbitrary singu-larity curve would decay faster than in a curveletframe. This rate of decay is much faster than thatof any other known system, including wavelets.Improved coefficient decay gives optimally sparserepresentations that are interesting in image-processing applications, where sparsity allows forbetter image reconstructions or coding algorithms.

Beyond Scale-Space?A beautiful thing about mathematical transformsis that they may be applied to a wide variety of prob-lems as long as they have a useful architecture. TheFourier transform, for example, is much more thana convenient tool for studying the heat equation(which motivated its development) and, by exten-sion, constant-coefficient partial differential equa-tions. The Fourier transform indeed suggests afundamentally new way of organizing informationas a superposition of frequency contributions, aconcept which is now part of our standard reper-toire. In a different direction, we mentioned beforethat wavelets have flourished because of their ability to describe transient features more accu-rately than classical expansions. Underlying thisphenomenon is a significant mathematical archi-tecture that proposes to decompose an object into a sum of contributions at different scales andlocations. This organization principle, sometimesreferred to as scale-space, has proved to be veryfruitful—at least as measured by the profound influence it bears on contemporary science.

Curvelets also exhibit an interesting architecturethat sets them apart from classical multiscale rep-resentations. Curvelets partition the frequencyplane into dyadic coronae and (unlike wavelets) subpartition those into angular wedges which again display the parabolic aspect ratio. Hence,the curvelet transform refines the scale-space view-point by adding an extra element, orientation, andoperates by measuring information about an object at specified scales and locations but onlyalong specified orientations. The specialist will rec-ognize the connection with ideas from microlocalanalysis. The joint localization in both space andfrequency allows us to think about curvelets as living inside “Heisenberg boxes” in phase-space,while the scale/location/orientation discretizationsuggests an associated tiling (or sampling) ofphase-space with those boxes. Because of this organization, curvelets can do things that other sys-tems cannot do. For example, they accurately modelthe geometry of wave propagation and, more gen-erally, the action of large classes of differentialequations: on the one hand they have enough frequency localization so that they approximatelybehave like waves, but on the other hand they have

enough spatial localization so that the flow will essentially preserve their shape.

Research in computational harmonic analysis involves the development of (1) innovative andfundamental mathematical tools, (2) fast compu-tational algorithms, and (3) their deployment in various scientific applications. This article essen-tially focused on the mathematical aspects of thecurvelet transform. Equally important is the sig-nificance of these ideas for practical applications.

Multiscale Geometric Analysis?Curvelets are new multiscale ideas for data repre-sentation, analysis, and synthesis which, from abroader viewpoint, suggest a new form of multiscaleanalysis combining ideas of geometry and multi-scale analysis. Of course, curvelets are by no meansthe only instances of this vision which perceivesthose promising links between geometry and mul-tiscale thinking. There is an emerging communityof mathematicians and scientists committed to the development of this field. In January 2003, for example, the Institute for Pure and Applied Mathe-matics at UCLA, newly funded by the National ScienceFoundation, held the first international workshopon this topic. The title of this conference: MultiscaleGeometric Analysis.

References[1] E. J. CANDÈS and L. DEMANET, Curvelets and Fourier in-

tegral operators, C. R. Math. Acad. Sci. Paris 336 (2003),395–398.

[2] E. J. CANDÈS and D. L. DONOHO, New tight frames ofcurvelets and optimal representations of objects withpiecewise C2 singularities, Comm. Pure Appl. Math.,to appear.

[3] H. F. SMITH, Wave equations with low regularity coef-ficients, Doc. Math., Extra Volume ICM 1998, II (1998),723–730.

Some curvelets at different scales.Candes 02, Do 02

Page 9: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

• Global & linear

• Similar to F-K Filtering

Contourlet Band Muting(Linear Filtering)

NoisyData

ContourletTransform

FrequencyBand Muting

Result

InverseTransform

Page 10: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

Contourlet Band FilteringSynthetic Example

50 100 150 200 250 300 350 400 450 500

50

100

150

200

250

300

350

400

450

500−1

−0.8

−0.6

−0.4

−0.2

0

0.2

0.4

0.6

0.8

1

Steep Parabolic Curves and Dispersive Ground Roll

Page 11: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

Contourlet Coefficient Sectors

Each band represents a group of coefficients that represents and individual part of the signal

Page 12: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

• Steep events removed

• Artifact located at apex

Reconstructed Contourlet Denoised Signal

Problems:Data Denoised by Contourlet Transform

50 100 150 200 250 300 350 400 450 500

50

100

150

200

250

300

350

400

450

500−0.25

−0.2

−0.15

−0.1

−0.05

0

0.05

0.1

0.15

0.2

0.25

Page 13: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

• Smart

• Local in Position and Dip

• Allows Incorporation of Prior Predictions

• Flexible

• Relatively Phase Insensitive

Curvelet Adaptive Subtraction

Page 14: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

Curvelet Adaptive Subtraction(Non-Linear Thresholding)

Ground RollEstimate

Threshold CurveletTransform

NoisyData

Result

CurveletTransform

InverseTransform

Page 15: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

Diagonal of the CovarianceCovariance

Curvelet Adaptive Subtractionnoisy data︷︸︸︷

d = m︸︷︷︸noise-free

+pred. noise︷︸︸︷

n

m̂ : minm12‖C−1/2

n (d−m) ‖22

ˆ̃m : minm̃

12‖Γ−1

(d̃− m̃

)‖22 + λ2‖m̃‖p

Cn = Γ2 =

Page 16: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

Curvelet Adaptive Subtraction

m̂ = BT ΘλΓ (Bd)

Γ = |Bnp|

Control ParameterHard or Soft ThresholdΘλΓ =

λ =

Page 17: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

Curvelet Adaptive Subtraction Synthetic Example

denoised

50 100 150 200 250 300 350 400 450 500

50

100

150

200

250

300

350

400

450

500−0.25

−0.2

−0.15

−0.1

−0.05

0

0.05

0.1

0.15

0.2

0.25

Page 18: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

Predicted NoisePredicted Noise

50 100 150 200 250 300 350 400 450 500

50

100

150

200

250

300

350

400

450

500−1

−0.8

−0.6

−0.4

−0.2

0

0.2

0.4

0.6

0.8

1

Page 19: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

Ground Roll Difference:Shown is the difference between the estimated ground

roll and the actual ground roll

Difference Between Modeled Ground Roll and Ground Roll Used in Synthetic

50 100 150 200 250

100

200

300

400

500

600

700

800

900

1000−1

−0.8

−0.6

−0.4

−0.2

0

0.2

0.4

0.6

0.8

1Difference Between Modeled Ground Roll and Ground Roll Used in Synthetic

50 100 150 200 250

100

200

300

400

500

600

700

800

900

1000−0.1

−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08

0.1

Difference is on the order of the ground roll

Page 20: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

Curvelet Adaptive

Subtraction: Real Data Example

200

400

600

800

1000

20 40 60 80

Oz25 Signal With Ground Roll

Page 21: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

Radon Predicted

Noise:200

400

600

800

1000

20 40 60 80

Two Choices:

1. Subtract predicted noise from data

2. Use predicted noise to define threshold for curvelet adaptive subtraction

Page 22: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

Low Quality Radon

Denoised Data

Subtraction of the Radon predicted noise

from the data

200

400

600

800

1000

20 40 60 80

Page 23: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

• Increased Smoothing

• Some removal of top reflectors and right side of mid reflectors

Curvelet Denoised Data From Soft

Non-Linear Threshold

200

400

600

800

1000

20 40 60 80

Page 24: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

• Better reflector preservation

• Less smoothing

Curvelet Denoised Data From Hard

Non-Linear Threshold

200

400

600

800

1000

20 40 60 80

Page 25: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

Curvelet Predicted Noise

200

400

600

800

1000

20 40 60 80

200

400

600

800

1000

20 40 60 80

Hard ThresholdingSoft Thresholding

Page 26: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

• Direct subtraction will no longer be useful as the result will only amplify the differences.

• Curvelet adaptive subtraction works without a problem.

Phase PreservationGive the predicted ground roll a 90 degrees phase shift. What would happen?

Page 27: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

Phase Shifted Model Results

200

400

600

800

1000

20 40 60 80

200

400

600

800

1000

20 40 60 80

Subtraction Curvelet Adaptive Subtraction

Page 28: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

Phase Shifted Model Results

Results Without Phase Shift Results With Phase Shift

200

400

600

800

1000

20 40 60 80

200

400

600

800

1000

20 40 60 80

Page 29: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

The Difference

200

400

600

800

1000

20 40 60 80

200

400

600

800

1000

20 40 60 80

200

400

600

800

1000

20 40 60 80

Results Without

Phase Shift

Results With Phase Shift

Difference

Page 30: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

Iterative Process

Curvelet AdaptiveSubtraction

Data1

Data2

Result1

Result2

Difference

NormalDifference

Phase ShiftedDifference

Curvelet AdaptiveSubtraction

Page 31: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

Iterative Result

200

400

600

800

1000

20 40 60 80

200

400

600

800

1000

20 40 60 80

Result After 3 Iterations Predicted Noise

Page 32: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

Iterative Result

200

400

600

800

1000

20 40 60 80

200

400

600

800

1000

20 40 60 80

0 Iterations 3 Iterations

Page 33: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

Iterative Effects

200

400

600

800

1000

20 40 60 80

10 20 30 40 50 60 70 80

100

200

300

400

500

600

700

800

900

1000−1

−0.8

−0.6

−0.4

−0.2

0

0.2

0.4

0.6

0.8

1

• Improves overall signal by removing artifacts though phase scanning

Difference Between Iterative Result and Initial Result

Page 34: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

High Quality Radon Comparison

200

400

600

800

1000

20 40 60 80

CAS Result High Quality Radon Result

Page 35: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

High Quality Radon Comparison

200

400

600

800

1000

20 40 60 80

CSA Noise High Quality Radon Noise

Page 36: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

• Curvelet Adaptive Subtraction can effectively remove Ground Roll and preserve reflectors

• Adaptive Subtraction works even with less than optimal noise modeling

• Curvelet flexibility allows for robust adaptive subtraction which is relatively phase independent

• Iterative process can further improve signal to noise ratio

Conclusions

Page 37: Imaging: Adaptive Ground Roll Removal · 2019-11-14 · number of di! erent ori entations. Using the second basis, the curvelet basis (5), we are able to better approximate the curve,

AcknowledgmentsWe would like to thank:

• Emmanuel Candes and David Donoho for the use of their codes.

• This work was in part financially supported by a NSERC Discovery Grant.

• Tadeusz Ulrych and the sponsors of the Consortium for the Development of Specialized Seismic Techniques (CDSST).