AASPI

Team
R&D GROUP, INTEG, GEOPIC
May 2024
GTM
Curvature Attributes

Team
R&D GROUP, INTEG, GEOPIC
May 2024
Curvature Attributes

4
Spectral Attributes

Team
R&D GROUP, INTEG, GEOPIC
May 2024
• Spectral attributes: To evaluate thin bed tuning, attenuation, and other
interference phenomena.
Spectral Continuous Wavelet Transform

7
Spectral Decomposition

• Spectral decomposition takes a broad band seismic signal and breaks it into a
finite number of spectral voice, magnitude, and phase components centered
about frequencies defined by the user
• Four different time-frequency decompositions based on matching pursuit,
continuous wavelet transform, maximum entropy (or least-squares spectral
analysis), and simple correlation.

8
CWT Spectral Magnitude components

9
10
Geometric Attributes

Team
R&D GROUP, INTEG, GEOPIC
May 2024
Geometric attributes

• Geometric attributes measure the orientation, continuity, and texture of 3D

seismic data that aid in mapping faults, folds, and flexures, as stratigraphic
features such as channels, mass transport complexes, and carbonate buildups
as well as drilling hazards.

12
 Dip3d

Team
R&D GROUP, INTEG, GEOPIC
May 2024
Dip Computation methods

1. Discrete Semblance Scans

2. Gradient structure Tensor

3. Ratio of smoothed instantaneous wavenumber to smoothed instantaneous

frequency

4. Trace to trace cross-correlation

5. Prediction error filter

Dip3d Parameterization

• Dip window height defines the half height of the analysis window, with a default
of 5 samples
• Smaller windows provide less stratigraphic mixing, whereas larger windows help
suppress random noise for the same wavelet.
• A good rule of thumb is to make the full window approximately equal to the
period of the dominant wavelet.
• Lateral window radius is bin size

15
Dip computation

16
Fault Mapping

Team
R&D GROUP, INTEG, GEOPIC
May 2024
Fault Mapping workflow
Fault Enhancement Workflow

1. Input Seismic attribute: Energy ratio similarity or

CNN Fault Probability

2. Inline dip

3. Crossline dip

4. Weight: Total Energy

Fault dip magnitude and dip azimuth

• Fault dip azimuth volume will exhibit

values ranging between ±180° and
the dip magnitude volume values
ranging between 0° and 90° for every
voxel.
• We will skeletonize these values by
setting them to be a znull value
wherever the fault probability falls
below the (5) Fault probability
threshold. If the Fault probability
threshold is set to 0, the output fault
dip azimuth and fault dip magnitude
volumes will not be skeletonized 20
 AASPI Fault Model

• The fault model’s input block size is 128 inlines x 128 crosslines x 128 vertical samples
• Overlapping window size defines the amount of overlap between adjacent blocks to be extracted from the input
volume and ranges between 0 and 128-1=127 voxels
Fault Enhancement

Fault enhancement attribute is a post-stack attribute

which enhances locally planar features within a
seismic attribute volume.
Will improve
1. Fault images
2. Enhance unconformities and other discontinuities
parallel or sub parallel to reflector dip
3. Enhance axial planes delineated by most-positive
and most-negative curvature volumes

22
• Window half-length

• Window half-width

• Window half-Height

window whose radius is three times the greater of these two spacing (XL and IL spacing)

Sigma1 and Sigma 3 define the width of Gaussian filters used in smoothing and sharpening respectively

Sigma 1

Sigma 2 (equals Sigma 1)

Sigma 3 (Should not be less than larger of cdp or inline spacing)

23
• Fault_enhancement pre computes the LoG operator as well as the attribute weighting matrix for azimuths
ranging between -180° and +180° and dip magnitudes ranging between 0° and 90°
• Initial testing indicates that there on several data volumes indicate no significant differences between sampling
at 1°, 3°, or 5°. The computation time is the same, independent as to how densely you store your operators.
• Let’s assume you use coherence as an input attribute to be enhanced. Internal to the program, the coherence,
c, is first converted to a fault probability, p=1-c, such that voxels exhibiting high coherence c≈1 are converted to
p≈0.0 . If all of the values of p in the analysis window are less than the (14) attribute minimum threshold, no
fault enhancement is attempted and the resulting fault probability is set to 0.0, thereby significantly reducing
the computation cost

24
• The parameters (15) θ1=Dip1 and (16) θ2=Dip2 define a Tukey filter that rejects fault attributes that fall beyond dip

magnitude θ1 and retains fault features beyond θ2. If the numerical value of θ1 > θ2, then fault features subparallel

to reflector dip are retained rather than suppressed. For example, if θ1= 10° and θ2=25° then all the discontinuities

with a dip lessthan 10° will be rejected, discontinuities with dip magnitude greater than 25° will be retained, and

discontinuities with dip magnitudes falling between 10° and θ2=25° will be suppressed using the filter described in

Figure 2. In contrast, if θ1=25° and θ2=10° then faults with a dip magnitude θ>25 0 will be rejected and

discontinuity features subparallel to reflector dip will be retained

• Numerical experimentation has shown that using a spherical window provides nearly the same result as (20) using

a rectangular prism window, but costs 6/π≈2 times less.

25
Fault Skeletonize

• Skeletonized3d is used to skeletonize discontinuous structures, such faults and

unconformities

Output of Fault Enhancement

Fault prediction using CNN

27
Input options

28
Dip3d results
Geometric attributes
• Coherence (or similarity) – measures waveform similarity. Different algorithms are sensitive only to changes in waveform shape
(energy ratio similarity) but not to changes in amplitude, or sensitive to lateral changes in amplitude as well (Sobel filter
similarity, semblance-based similarity, and chaos).
• Reflector dip – measure of the reflector orientation. Dip is a vector that can be represented in Cartesian coordinates using inline
and cross line apparent dip components or in polar coordinates using dip azimuth and dip magnitude.
• Reflector curvature – measures lateral change in dip. Curvature is represented by two vectors where the most useful are the
most-positive curvature, k1, and its strike ψ1, and the most-negative curvature, k2, and its strike ψ2.
• Reflector aberrancy (or flexure)– measures lateral change in curvature. Because it is the third derivative of time- or depth-
structure, aberrancy is represented by three vectors. Unlike most-positive and most-negative curvature, the minimum, middle,
and maximum aberrancy measures don’t correlate to easily recognized features; for this reason, we typically display the vector
sum of the three measures, the total aberrancy magnitude and total aberrancy azimuth.
• Reflector convergence –measure of the (for most relatively flat-lying strata) of the vertical change in vector dip. The
convergence vector is represented by an azimuth defining the direction of convergence (such as a pinch out or erosion
unconformity) and its magnitude.
Filter Bank

• Because tuning is frequency-dependent, we might anticipate that channel edges may appear more clearly on
some band limited versions of the data and less clearly on others.
• The default is a 50% taper such that f2= f3 for the filter banks, resulting in a suite of raised cosine tapers
• The original data spectral will be spanned by the Number of filter banks applied to the data.
Energy ratio similarity
Broadband vs Multispectral ERS
Amplitude Gradients

• Amplitude gradients are a measure of how the amplitude varies laterally along structural dip
 Coherent Energy

• Coherent energy (energy of the complex analysis window after Karhunen Loeve or principal component filtering)
 Stat3d

• Program stat3d computes the mean,

median, standard deviation, and user-
defined percentiles within structurally-
oriented windows
Gray Level Co-occurrence Matrix
(GLCM)
• Spectral attributes are used to evaluate thin bed tuning, attenuation, and other interference phenomena.

• Geometric attributes measure the orientation, continuity, and texture of 3D seismic data that aid in
mapping faults, folds, and flexures, as stratigraphic features such as channels, mass transport complexes,
and carbonate buildups as well as drilling hazards
Texture

• The classic definition of texture is to take your thumb (an analysis window!) and place it over a surface.
• As you feel the topographic asperities in this surface your brain may interpret the surface to be rough,
smooth, slick, ribbed, or waffled.
 Gray Level Co-occurrence Matrix

GLCM energy GLCM entropy GLCM homogeneity

How constant the reflector is within a How random the reflector is within a How smoothly varying the reflector is
window window within a window

• GLCM works on lateral “amplitude” asperities and will be sensitive to diagenetic alteration, chaotic or smooth
deformation, period fractures, or other features
Color Models

• Alpha blending, which we will see can be used to emulate a hue-saturation-lightness color model,

• red-green-blue (RGB), the color addition model, and

• cyan-magenta-yellow (CMY), the color subtraction model.

In practice, we need to worry about the “curse of dimensionality” (Bellman, 1957) where the volume of the N-
dimensional space increases so fast that all the data become sparse, thereby limiting the usefulness of any
statistical analysis.
• Principal component analysis (PCA) is widely used to reduce the redundancy and excess dimensionality of

the input attribute data

• Such reduction assumes that

1. most of the signal is represented by the first few principal components

2. random noise and anomalies are represented by the later principal components
Attributes

Identifying Depositional facies Characterizing geo-mechanical

variation variation in shale plays

1. dip magnitude 1. inversion volumes

2. coherence (similarity) 2. lambda-rho
3. GLCM texture attributes 3. mu-rho
4. spectral magnitude 4. intercept or gradient AVO volumes
5. coherent energy 5. anisotropy measure

Eigenvectors are linear features in N-dimensional attribute space and cannot represent cyclical attributes such
as phase, azimuth, or strike.
K1, K2 and Coherence

Plotting K1 and K2 corendered by ERS Plotting the curvature vector (The curvature strike
modulated by its value corendered by ERS)
• Aberrancy is a vector that measures the magnitude and orientation of flexures and is the 3rd derivative of an
implicit depth structure map.
Coherent Energy corendered with energy weighted amplitude gradient
 Emulating the HSL color model using alpha blending (opacity/transparency)

• Base layer should always be plotted against a polychromatic

color bar.
• The 2 layer will typically be plotted against a monochromatic
nd

gray color bar, which defines the saturation, or S, axis of the

HSL color model.

• The third axis will be plotted against a monochrome black color

bar for single polarity attributes and against a binary black-

white color bar for dual polarity attributes, which defines the

lightness, or L, axis of the HSL color model.

Relative Acoustic Impedance

Relative acoustic impedance is a simple trace integration scheme that approximates a band pass filtered version
of the absolute impedance when a zero-phase seismic wavelet can be approximated by a spike and when the
reflection coefficients are small

• Ormsby filter applied to the output volume to suppress possible low-frequency artifacts
Seismic Data conditioning types

1. Amplitude balancing: In general, amplitude balancing should be avoided, because it can negatively impact

any relative amplitude processing that has been applied to the data volume

2. Band-pass filtering

3. Spectral balancing: Goal of spectral balancing is to make the spectrum as flat as possible, without

increasing the noise

4. Structure-oriented filtering

5. Footprint suppression
Spec_cwt Parameterization
AGC

Long window AGC is safer than short window AGC in preserving relative amplitudes.
 RMS Amplitude

The root-mean-squared amplitude of the seismic data is one of the simplest and most used seismic attributes,
particularly when trying to map the energy within a package about a noisy horizon.

half-height of the RMS amplitude window, T=KΔT

 Voice Components

Attribute derived during spectral decomposition, called the voice components

Complex Wavelets used for CWT

Real and imaginary part of the complex wavelet

These complex wavelets are convolved with the real seismic traces to compute the complex spectral
components
Continuous Wavelet Transform
Peak Magnitude
Spectral decomposition in AASPI

In spectral decomposition, the seismic data are decomposed into their spectral components. The spectral
components can be expressed as the magnitude and phase, or as the voices.

AASPI has the following spectral decomposition algorithms:

1. Continuous wavelet transforms, spec_cwt, which runs the fastest,
2. Complex matching pursuit program spec_cmp, which runs slowest but has higher temporal and spectral
resolution
3. Constrained least-squares spectral analysis program spec_max_entropy, which also has very high temporal
and spatial resolution.
4. Spectral cross-correlation algorithm, program spectral_probe, which does not decompose an amplitude
volume, but rather generates a suite of normalized cross correlation coefficients with a suite of sine and
cosine single-cycle wavelets.
 Spectral decomposition

• Spectral decomposition takes a broad band seismic signal and breaks it into a finite number of spectral voice,
magnitude, and phase components centered about frequencies defined by the user.

• Spectral decomposition of seismic data helps in the analysis of subtle stratigraphic plays and fractured reservoirs
(Partyka et al., 1999; Marfurt and Kirlin, 2001)

1. Short-window discrete Fourier transform,

2. Continuous-wavelet transform (CWT),
3. S-transform, and
4. Matching-pursuit transform
1. Continuous wavelet transform algorithm: spec_cwt

2. complex matching pursuit: spec_cmp

3. maximum entropy spec_max_entropy

• Algorithms (2 & 3) are computationally more intensive but often provide improved spectral and temporal
resolution
When Spectral decomposition is carried out on seismic data, it yields the spectral magnitude and phase
at each time-frequency sample.

Spectral magnitude represents the energy that

correlates with the trace

Spectral phase represents phase rotation between the

seismic trace and the Morlet wavelet at each instant of
time

Voice component, is a simple function of spectral

magnitude, m, and phase φ at each time-frequency
sample

ν(t, f ) = m (t, f ) exp [-j φ(t, f)]

• For the continuous-wavelet transform, the voice
components are equivalent to narrow band pass filtered
versions of the input seismic data.
spectral-magnitude components have been used routinely to
delineate stratigraphic features at or below
the limits of seismic resolution
In this method, data are first decomposed into time-frequency spectral components.

Power of the spectral magnitude, P(t, f ) = m(t, f)^2, is averaged over all the traces (j = 1, …K) in the
data volume spatially and in the given time window, which yields a smoothed average power spectrum, given
by Pavg(t, f ).

Next, the peak of the average power spectrum, Ppeak(t), also is computed.

Both the average spectral magnitude and the peak of the average power spectrum are used to design a single
time-varying spectral-balancing operator that is applied to each and every trace in the data:

where ε is the prewhitening parameter. A conservative value

would be ε = 0.04.
Apparent dip

Using program apparent_cmpt, we can compute apparent dip at any arbitrary azimuth, φ, using the simple
formula:

papp(φ)= p cos(φ)+q sin(φ)

Multispectral Dip Components

• Hardage (2009) recognized that because of the variable signal-to-noise ratio at different frequencies, that
faults were more easily identified in his data on the low frequency components that were less contaminated
by strong inter-bed multiples.
Multispectral eigenvectors of the GST and dip components

Multispectral computation is based on the GST algorithm that measures the change of seismic amplitude (or
alternatively, the weighted change in the phase) in each of the three Cartesian directions.
Improving DIP estimates using filter_dip_components

An accurate estimate of dip is critical to the computation of geometric attributes

• Inline dip is an apparent dip along the inline axis

• Cross line dip is an apparent dip along the cross line axis

• Dip magnitude is displayed in degrees measured from the

horizontal.

• Dip azimuth ranges between -180° and +180° and is defined

clockwise from North, where North is at 0°, East at 90°, West
at -90°, and South at ±180°.
Filter Dip Components

1. LUM (lower-upper-middle),

2. MSMTM (multistage median-based modified trimmed mean) : Acts as an edge preserving filter, a

lineament preserving filter and can smoothen noise

3. Alpha-trimmed mean

4. Mean filter
Similarity

Coherence algorithm compares adjacent waveforms along

structure within an analysis window
 Structure-oriented filter (sof3d)

• Structure-oriented filtering provides a means

of preserving signal parallel to structural dip
while rejecting random and cross-cutting
coherent noise
• Similarity input control the edge preservation
components of the filters
Structure-oriented filter (sof3d)

• The principal component (also called Kohonen-Loève or KL) filter uses the data that fall within a 3D analysis
window defined by both lateral dimensions and a vertical analysis window.
Kuwahara Parameterization

• For similarity values s < slow, we assume we have a strong

edge, whereby the Kuwahara filtered data are assigned
weights of w=0.0

• If the value of the similarity attribute is greater than shigh <

s < scentered, and the Kuwahara filtered data are assigned
weights of w=1.0

• Finally, for values of similarity, slow < s < shigh we compute

a weight w=(s-slow)/(shigh-slow), and a compute the
linearly weighted average of the Kuwahara filtered and
unfiltered data dout=w*dfilt+(1-w)*dorig

Parameter selection such that preservation of • Threshold similarity value, s_center, above which the filter
low coherence geologic features of interest, always uses a centered analysis window.
suppress acquisition footprint, and enhance
more subtle geologic features.
Kuwahara filtering options

Four basic filters:

1. principal component filter,

2. alpha-trimmed mean filter: rejects outliers and smooths the remaining values

3. lower-upper-middle (LUM) filter: LUM filter preserves detailed variation, but rejects erroneous values.

4. mean filter

Note that a median filter is a subset of the alpha-trimmed mean filter when we set the rejected data on each

side to be 50%. If we set the rejected data to be 0%, we obtain the mean filter. The median filter is an edge-

preserving filter and will preserve changes in dips across faults. It also rejects erroneous spikes in the input data.
Techniques to preserve discontinuities across faults and stratigraphic edges:

1. Median filter along structural dip

2.
 Kuwahara Filter

• Kuwahara filter effects the data by both

 smoothening (by the application of the internal mean or other filter) and
 sharpening (by choosing a non-centered rather than centered window)

• Luo et al. (2002) were the first to apply a Kuwahara filter to seismic data.

• Luo examined the mean and standard deviation of a suite of overlapping windows of the same size that all
included the analysis point.

• Rather than replace the value of the analysis point with the mean of the centered window, the Kuwahara filter
replaces it with the mean of the (potentially non-centered) window that exhibits the smallest standard
deviation.
Kuwahara filter
Disorder

• Disorder algorithm is designed to emphasize noise and considers edges to be signal.

• Program disorder provides a good measure of the difficulty in mapping a horizon.

GLCM Entropy

GLCM entropy is sensitive to textural difference

 Structural Oriented filtering

Structure-oriented filtering provides a means of preserving

signal parallel to structural dip while rejecting random and
cross-cutting coherent noise.
 Spectral Decomposition

• Spectral decomposition methods can be divided into three classes:

1. Quadratic forms: Based on Wigner-ville distribution. Cannot be used for reconstruction

2. Linear forms: Based on Short time Fourier Transform (STFT) or CWT

3. Atom decomposition: Reconstructs the signal by using small “atom-sized” signals (in our case wavelets), such as matching

pursuit (program spec_cmp) and the Hilbert-Huang transform

spec_max_entropy is a nonlinear implementation of the
short time Fourier transform
 Matching Pursuit theory

• Spectra of the strongest events are estimated and

subtracted from the trace first, followed by iterative
estimation and subtraction of successively weaker
events.
Matching Pursuit algorithm
Matching Pursuit algorithm
The spectral bandwidth (the separation in Hz
between the 15th and 85th percentile of the
magnitude spectrum)
peak magnitude above average = peak magnitude - range-trimmed-mean spectrum
Spec_cmp Parameters

Smoothing window = 0.5 s

• Smoothes the spectra vertically before trying to estimate

a spectral balancing operator

Spectral balancing factor:

• 1%
• 4%, more conservative value and produce data closer to
original

For seismic data in the 5-120 Hz range indicates that a sample increment of Δf=2 Hz almost always gives good
reconstruction results.

Ricker wavelets tend to represent seismic data a little better, so convergence may be somewhat quicker even
though the resulting spectra will be very similar.

Vertical taper: A temporal taper is applied to ends of the input seismic data, thereby minimizing potential Gibb’s
phenomena.
Energy Ratio Similarity

• Numerically, energy ratio similarity is the ratio of the energy of the Karhunen-Loève filtered data over the
total (unfiltered) energy of the input data within the analysis window.
 Amplitude Volume Transform

AVT attribute (Bulhoes and Amorin) helps to

identify the different geologic features:

1. fault,

2. channel,

3. carbonate,

4. reflector unconformities and terminations

It is calculated by the root square of the average of the square of envelopes that are within a defined analysis window,
followed by the Hilbert transform
spectrally balanced amplitude from program spec_cmp or spec_cwt
K-means

• K-means takes multiple seismic attributes as input, and generates a facies volume
Attribute Selection

• For identifying the depositional facies variation, the volumetric attributes such as dip magnitude, coherency,
GLCM attributes, spectral magnitude, coherent energy can be considered as input

• For characterizing geo-mechanical variation in shale plays one should consider different volumes that help
in identifying the rock physics such as inversion volumes, lambda-rho, mu-rho, intercept or gradient AVO
volumes, etc.
Self Organizing Maps

• SOM is a projection from multidimensional attribute space to a 2D space

• Self-organizing map (SOM) is closely related to vector quantization methods (Haykin, 1999).

• SOM classification is initialized using the first two eigenvalues and eigenvectors. This 2D plane (the simplest
manifold in N-dimensional attribute space) is sampled by a suite of regularly spaced prototype vectors which
are then projected onto the SOM latent space.
Self Organizing Maps

• Assumption: Inputs are represented by J vectors in a N-dimensional vector space, where N is the number of
input attributes

• SOM consists of neurons or prototype vectors (PVs) organized by a lower-dimension grid, usually 2D, which
are representative of the input data that lies in the same N-dimensional space as the input seismic
attributes.

• PVs are also termed as SOM units and typically arranged in 2D hexagonal or rectangular structure maps that
preserve the neighborhood relationship among the PVs.

• In this manner PVs close to each other are associated with input seismic attribute vectors that are similar to
each other
SOM

• During the SOM training process, an input vector is initialized and is compared with all N-dimensional PVs
on the 2D grid, or latent space. The prototype vector with the best match (the winning PV) will be updated
as a part of SOM neighborhood training.
 SOM Algorithm

• Step 1: Consider an input vector, which is randomly chosen from the set of input vectors

• Step-2: Compute the Euclidean distance between this vector x and all PVs 𝐦i, i=1, 2,…p. The prototype vector 𝐦b, which

has the minimum distance to the input vector x, is defined to be the “winner” or the Best Matching Unit, 𝐦b

• Step-3: Update the “winner” prototype vector and its neighbors

• Step 4: Iterate through each learning step (steps 1-3) until the convergence criterion (which depends on the predefined

lowest neighborhood radius and the minimum distance between the PVs in the latent space) is reached

• Step 5: Color-code the trained PVs using 2D or 3D gradational colors (Matos et al. 2009)
End member index volume: Mahalanobis distance from a multi-attribute data vector to the center of all

prototype vectors, providing an estimate of the likelihood of a data vector being an end member (uncommon

facies).
Shortcomings of SOM

• There are two major limitations to the popular Self-organizing Maps (SOM) clustering algorithm.

1. There is no clear rule for selecting the training radius and the learning rate; these parameters are data

dependent.

2. Second, because of the absence of any defined cost function there is no measure of “confidence” in

the final clustering results to indicate the convergence at the final iteration.
GTM
a mass transport complex may be characterized by relatively low coherence, strongly converging reflectors, and
high entropy (measured by the GLCM algorithm)

Surrounding marine shales may be characterized by moderate coherence, low reflector convergence (i.e.,
parallel reflectors) and low GLCM entropy
Gray Level Co-occurrence Matrix (GLCM)

The classic definition of texture is to take your thumb (an analysis window) and place it over a surface. As you

feel the topographic asperities in this surface your brain may interpret the surface to be rough, smooth, slick,

ribbed, or waffled.

GLCM works on lateral “amplitude” asperities and will be sensitive to diagenetic alteration, chaotic or smooth

deformation, period fractures, or other features.

Unsupervised Learning

• “let the data speak for themselves” and find their own “natural” clusters that may or may not have anything
to do with our desired facies of interest
Mahalanobis Distance
KMeans Workflow

1. Generate training data

2. Analyze clustering algorithms

3. Create a k-means clustering model

4. Apply the k-means model to the entire input volumes

5. Display the results using cross plot or co-rendering.

KMeans Information

• We no longer output the cluster numbers, but rather their projections onto principal axes, because cluster
numbers do not tell us the distance between cluster centers.

• This allows us to know how similar or how close the cluster centers are to each other, using cross plot or RGB
co-rendering

• Data clipping percentile: To avoid the effect of extreme values on the plotting of histograms and partitioning
maps

• Clustering tolerance: Threshold of data fraction below which the iterative KMeans clustering process can stop
• Unlike k-means and Gaussian mixture models, SOM and GTM maintain a topological relationship between the
different clusters by projecting the data onto a deformed 2D manifold which in turn is mapped onto a 2D
latent space.
• This latent space is in turn mapped to a 2D continuous color template. Attribute vectors assigned to similar
clusters will exhibit a similar color
Spectral attributes

• Spectral attributes are used to evaluate thin bed tuning, attenuation, and other interference phenomena

• four different time-frequency decompositions based on matching pursuit, the continuous wavelet transform,
maximum entropy (or least-squares spectral analysis), and simple correlation.
Gaussian Mixture Model (GMM)

• GMM attempts to model the probability density of a set of data and can also act as a clustering technique to

generate unsupervised seismic facies

• Gaussian mixture models can be considered a probabilistic formulation of the k-means algorithm where the

means of each cluster are also given a covariance.

• Model based classification comes with a couple assumptions.

• Dataset comes from a mixture of multivariate Gaussian distributions. In other words, the clusters are

assumed to be shaped like a multivariate Gaussian distribution.

• Classification approach implies that a voxel is produced by a single cluster, or Gaussian distribution.
 Gaussian Mixture Model (GMM)

• Mixture model parameters contains the means, covariances, and weights of the mixture model.

• Include cluster decomposition volumes outputs the posterior probability of each component as a different
seismic volume

• Neighborhood expectation-maximization (NEM) algorithm allows for spatial information to be included in the
optimization process.

• Neighborhood coefficient is a scalar value that gives weight to the spatial information;

• Setting the neighborhood coefficient to zero will result in conventional expectation-maximization algorithm
GMM

• More parsimonious models on the left and the more complex models on the right

• Minimum width of Gaussian is directly related to the covariance matrix and must be non-zero to prevent
covariance matrices from becoming singular