ipam mga tutorial: multiscale problems in...
TRANSCRIPT
IPAM MGA Tutorial: MultiscaleProblems in Geophysics
Naoki [email protected]
http://www.math.ucdavis.edu/˜saito/
Department of Mathematics
University of California, Davis
Sep. 2004 – p.1
Acknowledgment
Nick Bennett (Schlumberger)
Bob Burridge (MIT)
Charlie Flaum (Schlumberger)
T. S. Ramakrishnan (Schlumberger)
Anthony Vassiliou (GeoEnergy)
Sep. 2004 – p.2
Outline
Multiscale Nature of Geophysical Measurements
Importance of Geometry, Textures, and Clustering:High Relevance to MGA2004
Large Scale: Seismic Signal/Noise Separation
Small Scale: Fracture and Vug Detection fromElectrical Conductivity Images
High-Dimensional Clustering Problem in ExplorationGeophysics
Summary
Sep. 2004 – p.3
Multiscale Nature of Geophysical
Measurements
Large Scale: Surface Seismic Survey:
used for structural mapping of the entire reservoirvolume
wide spatial coverage: � 10km x 10km x 3km
spatial sampling: 25m x 25m
can resolve features > 30-40ft (9-12m)
vertical axis: time
Sep. 2004 – p.4
Surface Seismic Survey
Sep. 2004 – p.5
Multiscale Nature of Geophysical
Measurements . . .
Medium Scale: Borehole Seismics:
receivers are in a borehole/well
can record not only reflected waves but also directwaves
can record reflected waves with less attenuation
used for linking surface seismic information and welllog information (e.g., sonic logs)
depth sampling: 20-50m (wish: < 10m)
Sep. 2004 – p.6
Borehole Seismics
courtesy: Schlumberger Sep. 2004 – p.7
Borehole Seismics . . .
courtesy: Schlumberger
Sep. 2004 – p.8
Multiscale Nature of Geophysical
Measurements . . .
Small Scale: Wireline measurements, cores:
used for detailed geological/geophysical informationin the vicinity of wells
spatial coverage: � 1m x 1m x 3km
can resolve features � 0.5-3ft (0.15-1m) or less
vertical axis: depth
cores provide absolute truth, but expensive
can characterize not only elastic but also electricaland nuclear properties of subsurface formations
Sep. 2004 – p.9
Electrical Conductivity Image ( � 2.5ftx 2.3ft)
courtesy: Schlumberger Sep. 2004 – p.10
Electrical Conductivity Imaging Tool
����� � ��� � � ��� � � ��
�� ��� � ��� � � ���� � �� � � �� � �
FMI measurement current path.
(a) Tool (b) Pad/Flap
courtesy: Schlumberger Sep. 2004 – p.11
Multiscale Nature of Geophysical
Measurements . . .
Large Scale Geometry
Medium/Small Scale Geometry
Textures (Stratigraphy, Lithology, Facies)
Reconciliation of multiscale info for better models
Highly Relevant to MGA 2004
This tutorial: introduction of problems + some attemptswith older techniques
My hope: provoke the audience so that some of youmay be interested in applying new techniquesdiscussed in MGA 2004 to these problems!
Sep. 2004 – p.12
Extraction of Large Scale Geometric Objects
Harlan, Claerbout, & Rocca (1984)
Motivation: Separate geologic events (layerboundaries) from other events and noise. Other eventscould be interference patterns, scattering from faults inthe form of hyperbolas, etc.
Measurement: Seismic signals (large scale)
Beginning of “multilayered transforms,” “coherentstructure extraction,” and ICA
Method: iteration of “focusing” transforms + softthresholding
Sep. 2004 – p.13
Harlan, Claerbout, & Rocca (1984) . . .
Define “focusing” as an invertible linear transformationmaking the data more statistically independent (or lessstatistically dependent)
A transform focusing signal (pattern) must alsodefocus noise
Focused signal becomes more non-Gaussian;defocused noise becomes more Gaussian
In the transformed domain, apply filtering orsoft-thresholding operations to extract signal orseparate signal from noise
Sep. 2004 – p.14
Harlan, Claerbout, & Rocca (1984) . . .
A stacked seismic section = of
Geologic component � linear events
Diffraction component � hyperbolic events
Noise component � white Gaussian noise ��� .
Sep. 2004 – p.15
Radon Transform
The 2D-line version for
� ����� � � can be stated as follows:
"! �$#� % � & � ����� ' �( � '*) # ) %� �+ � + '-,
Has a huge list of applications (X-ray tomography,geophysics, wave propagation, . . . )
Can be generalized: integration of an .-dimensionalfunction over /-dimensional geometric objects
Inversion formulas exist as Beylkin showed yesterday
Lines in an image are transformed to sharp peaks in�$#� % � -domainSep. 2004 – p.16
Example 1
(a) Original (b) 021 3
(c) Randomized (d) 01 3
Sep. 2004 – p.17
Example 1 . . .
(e) Thresholded (f) Recon
(g) Residual (h) Hyp. Ext.Sep. 2004 – p.18
Example 2: Original
Sep. 2004 – p.19
Example 2: Reconstruction
Sep. 2004 – p.20
Example 2: Residual
Sep. 2004 – p.21
Challenges
Fast Radon and Generalized Radon Transforms & 4
Beylkin’s Fast Radon Transforms and USFFT
Apply curvelets/ridgelets & 4 F. Herrmann
3D is a big issue
Sep. 2004 – p.22
Fracture Plane Detection fromBorehole Images
Objective: detect and characterize fracture planesstriking a borehole
Measurement: borehole images (either electric oracoustic), medium 5small scale
Method: Hough transform
Fracture plane can be parameterized by6� � 7 � ��8 & 9,
The name of the game is to estimate the parameters� 6� 7� 9 �from available borehole images.
Sep. 2004 – p.23
Fracture Plane Cutting a Borehole
RO
x
y
z
(o,o,d)φ
α
(xi,yi,zi)x
x
ηi
θi
Sep. 2004 – p.24
A Synthetic Borehole Image(Unfolded)
0 100 200 300
050
100
150
200
250
300
Sep. 2004 – p.25
Detected Edges
0 100 200 300
050
100
150
200
250
300
Sep. 2004 – p.26
An Edge Element vs FractureGeometry
An edge location
���2: � �: � 8 : �
and an edge orientation ;:
constrain the fracture plane:
6�: � 7 �: ��8 : & 9
6 �: ) 7�: & <>= ? ;: ,
These two equations specify a straight line in the� 6� 7� 9 �
-space:
6 & 6 � 9 � & 9) 8 : @ �: � �: <>= ? ;:
7 & 7 � 9 � & 9) 8 : @ �: ) �: <= ? ;: ,
Sep. 2004 – p.27
Hough Transform
is a popular technique to detect certain geometric patternsfrom the edges in images. Informally, it is related to Radontransform. For line detection in 2D:
A � 6� 7 � & B �CD E � ����� � � C �( � 7 � 6� ) � �+ � + ��
where
B
is a thresholding operation to make a binary image,D E is a regularized derivative with a characteristic scale F.But the HT more heavily utilizes the duality between the fea-
ture space and the parameter space.
Sep. 2004 – p.28
Hough Transform . . .
Hough (1962) patented the original method for binaryimages
Ballard (1981) generalized to arbitrary shape withgradient info
Illingworth & Kittler (1988) surveyed and listed 136papers
Leavers (1993) surveyed and listed 173 papers
Sep. 2004 – p.29
Hough Transform . . .
Compute edge positions and orientations (gradient)
Prepare the parameter space (called accumulatorarray) whose axes are the parameters specifyingshapes (e.g.,lines: gradient and �-intercept, circles:center and radius)
For each edge, vote for all possible (specific) shapes inthe accumulator array
Accumulate the votes for all (significant) edges
Local maxima in the accumulator array identifyconcrete shapes
Sep. 2004 – p.30
Hough Transform . . .
+ Very robust, insensitive to broken patterns andnoises
+ Can detect multiple or intersecting objects
+ Can be more effective given edge orientationinformation
- Computationally slow (voting process)
- High storage (memory) requirement
Discretization of parameter space G templatematching in feature space
Sep. 2004 – p.31
Edge Orientation ConstrainsPossible Planes
0 100 200 300
050
100
150
200
250
300
Sep. 2004 – p.32
Detected Fracture Planes
0 50 100 150
050
100
150
200
250
300
Sep. 2004 – p.33
Real Data Result
courtesy: Schlumberger Sep. 2004 – p.34
Detection of Vugs/Elliptic Shapes
Objective: detect and characterize vugs (ellipticalcavity or void in a rock) & 4 porosity, depositionalinformation
Measurement: borehole images (electric or acoustic),small scale
Method: Hough transform for ellipses
Sep. 2004 – p.35
Hough Transform for Ellipses
5 parameters (center coordinate
����� � � , lengths ofmajor/minor axes
� 6� 7 �
, and orientation of major axis
H
)are required to specify each ellipse
Voting in the 5 dimensional accumulation array is costly
Found lower dimensional strategy if one needs onlycertain combinations of parameters such as centerpositions and areas of ellipses (N. Bennett, R.Burridge, & NS)
Sep. 2004 – p.36
Hough Transform for Ellipses . . .
Each pair of edges in an image determines aone-parameter family of ellipses (an exercise ofprojective geometry).
Let
IKJ & ��� J� � J � and L J & � % J� M J � , N & O� P
be thepositions and the normal vectors of a pair of edges.
Sep. 2004 – p.37
Hough Transform for Ellipses . . .
Then, the following represents a conic section
Q ����� � � :
Q ����� �R S � T & U @ ����� � � ) S VXW ����� � � V @ ���� � � & Y�
where
U ����� � � T &Z[Z[Z\Z[Z[Z[Z[Z� � O
� W �W O
� @ � @ OZ[Z[Z\Z[Z[Z[Z[Z
& Y
is the lineIW I @ ,
V J ����� � � T & % J �� ) � J � � M J � � ) � J � & Y
is the tangent line at
I J . Sep. 2004 – p.38
Hough Transform for Ellipses . . .
If
S^] � Y� S: �
, then
Q ����� �R S �
represents an ellipsewhere
S: & _ � L W ` ) ) aIW I @ � � L @ ` ) ) aI @ IW �b � %^W M @ ) % @ MW � @ ,
Other cases:
Sdc Y
orSde S: : hyperbola;
S & Y
: lineIW I @ ; S & S: : parabola.
All the geometric quantities of interest (e.g., all those 5parameters as well as its area: f 6 7) can be written as afunction of
S.
Sep. 2004 – p.39
Hough Transform for Ellipses . . .
0 1 2 3 4 5 6 7 8 9 100
1
2
3
4
5
6
7
8
9
10
Sep. 2004 – p.40
Hough Transform for Ellipses . . .
Position and area of ellipses are of particular interest
Need to consider only
g h
curve���: � S �� �: � S �� f 6 � S � 7 � S � �
instead of resolving 5parameters
For each pair of all the strong edges (or its randomsubset), vote for all ellipses lying along this curve
Pick the local maxima in this 3D voting space(accumulator array)
Sep. 2004 – p.41
Five Ellipses
0.0 0.2 0.4 0.6 0.8 1.0
10.
80.
60.
40.
20
(a) Original
0.0 0.2 0.4 0.6 0.8 1.0
10.
80.
60.
40.
20
(b) Edge Mag.
0.0 0.2 0.4 0.6 0.8 1.0
10.
80.
60.
40.
20
(c) Edge Ori.
0.0 0.2 0.4 0.6 0.8 1.0
10.
80.
60.
40.
20
(d) Detected Cnt. Sep. 2004 – p.42
Five Ellipses: Orientation vs Area
-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5
0.0
0.2
0.4
0.6
0.8
1.0
ORIENTATION
AR
EA
Sep. 2004 – p.43
Challenges
Fast Hough Transform via multiscale?
Will beamlets or its generalized version help?
Other geometric shapes?
Remember that there are strong needs in manyapplications to specific geometric shapes from images!
Sep. 2004 – p.44
A Library of FocusingTransformations
Each focusing transformation focuses a specificgeometric object
Can repeat extraction of different objects sequentially
A Library of Focusing Transformationsh 6 ' 6 & i 7jlk m 'on W � i 7jlk m 'on @ � ` ` ` ,e.g., A borehole image = fractures + vugs + residuals.
Sep. 2004 – p.45
Reconciliation of Different Scales
Small & 4Large: Geometric information from smallerscale measurements serves as constraints for modelsof seismic imaging and inversion
Large & 4Small: Geometry from surface seismic helpswell-to-well correlation (for avoiding “local minima”)
Easier said than done
Sep. 2004 – p.46
Reconciliation of Different Scales . . .
courtesy: Schlumberger Sep. 2004 – p.47
Clustering of Multiscale GeophysicalInformation
Objective: identify clusters of various attributes ofmeasurements; associate them with the facies (rocktypes); compare them with cores
Measurements include:
The standard well logs (porosity, density, . . . )
Statistics (mean, variance, skewness, kurtosis) ofborehole images (electric) with sliding windows;
Wavelet features of borehole images (electric) withsliding windows
Method: Self-Organizing Map (SOM) algorithm
Sep. 2004 – p.48
Wavelet-Based Texture Features
Decompose available images into local frequencycomponents via wavelet packets
Select the best basis that captures majority of energyof images
Represent images based on the selected basisfunctions
Supply the square of the top
p
(say,
O Y Y
) coefficients toSOM
Sep. 2004 – p.49
FaciesC
lassificationfrom
Wavelet
Features
Depth
04*10^7
-9170 -9150 -9130 -9110 -9090-9170 -9150 -9130 -9110 -9090
Depth
04*10^5
-9170 -9150 -9130 -9110 -9090-9170 -9150 -9130 -9110 -9090
Depth
010^5
-9170 -9150 -9130 -9110 -9090-9170 -9150 -9130 -9110 -9090
Depth
04*10^4
10^5
-9170 -9150 -9130 -9110 -9090-9170 -9150 -9130 -9110 -9090
Sep.2004
–p.50
FaciesC
lassificationfrom
Wavelet
Features...
Depth
02*10^4
5*10^4
-9170 -9150 -9130 -9110 -9090-9170 -9150 -9130 -9110 -9090
Depth
010^4
-9170 -9150 -9130 -9110 -9090-9170 -9150 -9130 -9110 -9090
Depth
04*10^3
10^4
-9170 -9150 -9130 -9110 -9090-9170 -9150 -9130 -9110 -9090
Depth
05*10^3
-9170 -9150 -9130 -9110 -9090-9170 -9150 -9130 -9110 -9090
Sep.2004
–p.51
Self-Organizing Maps (SOM)
A biologically motivated clustering method proposed byT. Kohonen ( 5 1981)
Viewed as a nonlinear projection of the probabilitydensity function % �rq �
of high dimensional input vectorq ] s t
onto the 2D plane
Neighboring points ins t
are mapped to neighboringpoints in
s @Also viewed as a vector quantization: approximate anyinput vector by the closest vector in a set of referencevectors
The Euclidean distance is often used. Sep. 2004 – p.52
Self-Organizing Maps (SOM) . . .
Consider a 2D array of nodes organized as ahexagonal lattice.
At a node located at u J , an initial reference (random)vector v J � Y � ] s t
is associated,
N & O� , , ,� w
.
A sequence of input vectors q � ' � ] s t
,
' & Y� O� , , , arecompared with the reference vectors
x v J � ' �y
.
For each
'
, the “best-matching” node
� u{z� vz � ' � �
withq � ' �
is found:m &= | } ~ � ?W� J� � � q � ' � ) v J � ' � ��
� q � ' � ) vz � & ~ � ?W� J� � � q � ' � ) v J � ' � � ,
Sep. 2004 – p.53
Self-Organizing Maps (SOM) . . .
Then, the neighboring nodes are updated by thefollowing rule (learning process):
v J � ' � O � & v J � ' � � �z J � ' �� q � ' � ) v J � ' ���
where
�z J � ' �
is called the neighborhood function, anon-negative smoothing kernel over the lattice points.
Normally,
�z J � ' � & � � � uz ) u J �� ' �
with
� � `� ' �� Y
as' � �, � ���� ` �� Yas � � �.
Sep. 2004 – p.54
Self-Organizing Maps (SOM) . . .
Typical examples of
�z J � ' �
:
�z J � ' � & � � ' ���� � �) � uz ) u J � @b P�� @ � ' � ��
or �z J � ' � & � � ' ��� �$� �� � � N ��
where� � ' �
is a learning-rate factor
Y c � � ' � c O
, � � ' �
isan effective width of the kernel, and
wz � ' �
is aneighborhood set of the best node m, all of which arenon-increasing functions of
'
.
Plasticity . . .
Sep. 2004 – p.55
Self-Organizing Maps (SOM) . . .
Sep. 2004 – p.56
Self-Organizing Maps (SOM): AnExample
Sep. 2004 – p.57
Facies Classification from WaveletFeatures
Feature vectors where the ground truth of facies wereobtained from thin sections were fed to the trainedSOM, which resulted in the labels (G, P, M, W).
Clusters corresponding Grainstone-Packstone (grainsupported) and Mudstone-Wackstone (mud supported)were identified.
G G G
P G P M W
P G G W W M
P W M Sep. 2004 – p.58
Facies Class
courtesy: Bureau of Economic Geology: UT AustinSep. 2004 – p.59
Remarks on SOM
- Many parameters to be specified, and theperformance critically depends on some of them.
- Precise mathematical analysis of convergence istough (proof done only for 1D array of nodes).
+ Each iteration is computationally fast.
Sep. 2004 – p.60
Challenges
Apply “Diffusion Maps” to the data!
Issue of normalization of different measurements
Separation of environmental effects and truegeophysical properties of rocks and formations frommeasurements (the uncertainty principle!)
Sep. 2004 – p.61
Summary
Reviewed three geophysical problems
Detection and description of geometry followed bycharacterization of texture
Scale of Measurements: vast range
Dimensions of Measurements: high
Classical techniques have been used
Can significantly improve via new techniquesdiscussed in this IPAM MGA program
May be able to provide data if interested
Sep. 2004 – p.62
References
W. S. Harlan, J. F. Claerbout, and F. Rocca: “Signal/noiseseparation and velocity estimation,” Geophysics, vol.49,pp.1869–1880, 1984.
N. N. Bennett, R. Burridge, and N. Saito, “A method to detect andcharacterize ellipses using the Hough transform,” IEEE Trans.Pattern Anal. Machine Intell., vol. 21, no. 7, pp.652–657, 1999.
N. Saito, N. N. Bennett, and R. Burridge, “Methods of DeterminingDips and Azimuths of Fractures from Borehole Images,” USPatent Number 5,960,371, Grant Date: 9/28/99.
N. Saito, A. Rabaute, and T. S. Ramakrishnan, “Method forInterpreting Petrophysical Data,” UK Patent Number GB2345776,Grant Date: 1/16/01.
Sep. 2004 – p.63
References on Radon Transforms
G.Beylkin: “Discrete Radon Transform,” IEEE Trans. Acoust.,Speech, Signal Processing, vol.ASSP-35, no.2, pp.162–172,1987.
S.R.Deans: The Radon Transform and Some of Its Applications,Wiley Interscience, 1983.
I.M.Gel’fand, M.I.Graev, and N.Ya.Vilenkin: GeneralizedFunctions, vol.5, Integral Geometry and Representation Theory,Academic Press, 1966.
L.A.Santaló: Integral Geometry and Geometric Probability,Addison-Wesley, 1976.
L.A.Shepp and J.B.Kruskal: “Computerized tomography: The newmedical x-ray technology,” Amer. Math. Monthly, vol.85,pp.402–439, 1978.
Sep. 2004 – p.64
References on Hough Transform
J.Illingworth and J.Kittler: “A survey of the Hough transform,”Comput. Vision, Graphics, Image Processing, vol.44, pp.87–116,1988,
V.F.Leavers: “Which Hough transform?” CVGIP: ImageUnderstanding, vol.58, pp.250–264, 1993, This article cites 173articles!
J.Princen, J.Illingworth, J.Kittler: “A formal definition of the Houghtransform: Properties and relationships,” J. Math. Imaging andVision, vol.1, pp.153–168, 1992.
Sep. 2004 – p.65
References on SOM
T. Kohonen: Self-Organizing Maps, Springer-Verlag, 1995.
Sep. 2004 – p.66