The chimney cube, an example of semi-automated detection of seismic objects by directive
attributes and neural networks: Part I; methodology
Paul Meldahl* and Roar Heggland, Statoil, Bert Bril and Paul de Groot, De Groot-Bril Earth Sciences
A new method for semi-automated detection of seismic
objects and interfaces such as faults, reflectors and
chimneys is presented.
The method increases the detectability and mapping
efficiency of the desired objects by an iterative process
comprising at least two steps: contrasting (i.e. texture
enhancement) followed by object recognition.
Contrasting is performed by extracting several attributes
from multiple windows and feeding these to either a
supervised, or an unsupervised neural network. The size,
shape and direction of the extraction windows as well as
the attributes are chosen in relation to the objects we wish
to detect. The windows may have a fixed shape and
direction, or they have data adaptive forms. In the latter
case they follow the local dip and azimuth of the seismic
events (Fig. 2). The resulting output is a texture enhanced
volume, which can be interpreted manually, or used as
input to the object recognition phase. This part is still under
development hence we will focus in this paper on the
texture enhancement phase.
In part I the method is explained and potential applications
in image processsing, structural and litho-stratigraphic
interpretation and reservoir characterization are discussed.
The chimney cube, a seismic volume, which highlights
vertical disturbances is introduced and it is discussed how
this volume was created.
In part II (Heggland et al. 1999) of this sequel
interpretations of chimney cubes are presented. For
example, it will be shown that the chimney cube may add a
new dimension to seismic interpretation as an indirect
hydrocarbon indicator tool (Fig. 1).
Seismic attributes and supervised and unsupervised neural
networks have become increasingly popular in recent years
in the realm of quantitative interpretation. In this paper we
extend the use of these techniques to seismic object
detection. Moreover, we introduce the concept of
directivity in the attribute extraction process.
Directive seismic source arrays have been used for many
years to attenuate unwanted signals hence increasing the
contrast between desired and unwanted energy. Since
seismic acquisition must record all desired energy the
source directivity is generally weak. Also in processing the
concept of directivity is used to increase the contrast
between objects and their background. Also these
directivity processes are weak since they should not
attenuate energy from seismic objects of interest.
In this method we improve seismic object detection by:
focusing on one class of objects only
using directivity to extract the attributes
the use of neural networks to recombine the extracted
attributes into new attributes with improved separation
The target can be reflections, faults, chimneys, seismic
anomalies or any other object of interest. The seismic
texture, the spatial extension and orientation of each of
these objects is different. Differences are both due to the
seismic response and how the data has been handled in
acquisition and processing.
Fig. 1 extracted chimneys and a time horizon
The chimney cube: Part I; methodology
Statement of the problem
To detect seismic objects requires knowledge about texture,
size, shape and direction of the objects. We have to ask
ourselves what is characteristic of a fault, chimney or
seismic anomaly in order to extract the best attributes.
These attributes are then recombined into even better
attributes via neural network mapping so that the objects
can be detected in an optimal way.
For example, faults are in general dipping more steeply
then reflectors and the seismic response changes faster
along fault planes than along reflectors. Since fast spatial
variations are mostly degraded by inaccuracies in
acquisition and processing we know that reflectors usually
contain higher temporal frequencies than fault images.
Seismic chimneys on the other hand appear as vertically
degraded zones in the seismic image. These zones can
completely mask the reflection energy from the
sedimentary sequence. (Heggland, 1997). We will come
back to these characteristics later in this paper when we
discuss how this knowledge was utilized in the processing
of the chimney cube.
Other examples of seismic objects and their characteristics
are: Direct Hydrocarbon Indicators (DHI)) and stratigraphic
units. A DHI is a seismic anomaly, which is often
characterized by a horizontal component, a change in
amplitude and phase and a termination against other
reflectors. A stratigraphic unit can have many different
responses. Usually the response changes along the
reflecting unit, due to changes in rock and fluid parameters.
Detecting these changes and relating these to geological /
petrophysical variations is the subject of seismic reservoir
characterization (e.g. de Groot, 1999a). However, if the
general response of a particular unit differs from the
surrounding reflectors, this information can be used in an
alternative auto-tracking scheme.
Once the decision is made which objects we wish to detect
we make an intelligent selection of attributes that have
potential to increase the contrast. Attributes can be
amplitude, energy, similarity, frequency, phase, dip,
azimuth etc. Moreover, attributes can be extracted (and
merged) from different input cubes e.g. near - and far offset
stack, inverted Acoustic Impedance etc. The attributes are
made directive by the shape and orientation of the
extraction window, (actually an extraction volume, see Fig.
2). In the chimney detection example we used three
vertically oriented extraction volumes to reflect that we are
looking for vertically oriented bodies of considerable
dimensions. We use knowledge about the characteristics of
chimneys by calculating in each extraction volume such
attributes as energy and various types of trace-to-trace
In fault detection, static, vertically oriented calculation
volumes can also be used. To prevent non-vertical faults
from “falling out of” the extraction volume(s) the vertical
directivity can be reduced. Reducing the vertical extension
and increasing the horizontal extension of the extraction
volumes does this.
To detect reflectors the calculation volumes may be
oriented horizontally. Again since reflectors are not
perfectly horizontal the directivity may be reduced.
So far we assumed the extraction volumes to be either
cubes or cylinders. Other forms may perform better,
especially in the case where the objects do not have a fixed
direction. For example, to detect faults we expect energy to
be an important attribute. In the ideal case we want to
calculate the energy in a 2D window along the fault plane.
As the orientation of the fault plane is unknown we must
reduce the directivity e.g. by using a cone shape extraction
volume to compute the energy attribute.
The ideal extraction volume follows the desired object at
every position. This implies that the extraction volume has
a flexible shape, which follows the local dip and azimuth of
the data (Fig. 2). The local dip and azimuth can be
calculated in many different ways. We use a modified
version of the Wigner-Radon transformation scheme
developed by Steeghs (1997). We found out that the
calculated local dip and azimuth cannot only be used to
steer the attribute extraction volumes but it is also a perfect
vehicle to remove random noise prior to attribute extraction
Fig. 2 Principle of directivity in attribute extraction
windows (the figure is drawn in two dimensions, in
reality the windows are flexible in three dimensions,
hence we are actually dealing with extraction volumes).
The chimney cube: Part I; methodology
After the selected attributes have been extracted at a
representative set of data points we will recombine these
into a new set of attributes to facilitate the detection
process. In this step we use supervised or unsupervised
The main difference between supervised and unsupervised
learning approaches lies in the amount of a-priori
information that is supplied. Supervised learning requires a
representative set of examples to train the neural network.
For example networks can be trained to find the (possible
non-linear) relation between seismic response and rock
property of interest (e.g. de Groot, 1999a and b). In this
case the training set is constructed from real or simulated
well data. In unsupervised (or competitive learning)
approaches, the aim is to find structure within the data and
thus extract relevant properties, or features. The resulting
data segments (patterns) still need to be interpreted. An
example of this approach is the popular waveform
segmentation method whereby waveforms along an
interpreted horizon are segmented. The resulting patterns
are then interpreted in terms of facies- or fluid changes.
In the object detection method we use the same principles.
If we employ unsupervised learning approaches we use
attributes related to the objects we would like to detect.
With supervised learning approach we go one step further.
Not only do we use meaningful attributes. We also identify
locations in the seismic cube where examples of the class
of objects to be detected are present. Seismic attributes are
calculated at these positions as well as at control points
outside the objects. The neural network is then trained to
classify the input location as falling inside or outside the
object. Application of the trained network yields the
desired texture enhanced volume in which the desired
objects can be detected more easily.
Edge detection algorithms and pattern recognition tools can
now be applied to the texture enhanced volume to further
improve the detectability of the objects. The concept of
directivity can also be applied in these processes.
Application: the chimney cube
We introduce a new seismic entity, which we call the
chimney cube. A chimney cube is a 3D volume of seismic
data, which highlights vertical disturbances of seismic
signals. These disturbances are often associated with gas
chimneys. The cube facilitates the difficult task of manual
interpretation of gas chimneys. It reveals information on
the hydrocarbon history and fluid flow models. In other
words the chimney cube may reveal where hydrocarbons
originated, how they migrated into a prospect and how they
spilled from this prospect. As such a chimney cube can be
seen as a new indirect hydrocarbon indicator tool.
Chimney interpretation is also used in geo hazard
evaluation. Correlating chimneys with mapped shallow gas
indicators may confirm the presence of shallow gas.
As chimneys are signs of partially degraded data, the cube
can also be used as a quality control tool in processing and
in the evaluation of attribute and depth maps.
Finally we see applications of the cube in determining
acquisition parameters. For example the success of 4C
seismic depends on our ability to undershoot gas, hence it
depends on the interpretation of chimneys.
The chimney cube whose interpretation will be the subject
of part II of this sequel was created as follows:
1. A seed interpretation was made with locations inside
manually interpreted chimneys and in a control set
outside the chimneys (Fig. 3).
2. At the seed locations various energy and similarity
attributes were extracted in three vertically aligned
extraction volumes around the locations (directivity
3. Step 1 and 2 were repeated to create an independent
4. A fully connected Multi-Layer-Perceptron type of
neural network was trained to classify the attributes
into two classes representing chimney or non-chimney
(output vectors 1,0 or 0,1). Fig. 4 shows the network
5. The trained network was applied to the entire data set
yielding outputs at each sample location. As the
outputs are complementary we passed only the output
on the chimney node to produce the final result: a cube
with values between approx. 0 (no-chimney) and 1
(chimney), see Fig. 5.
A new semi-automated method on detection of seismic
objects was presented. The method, which has wide
applicability in seismic processing and interpretation is
1. Focussing on one class of objects at the time
2. Extraction of attributes with potential to increase the
contrast between desired objects and the background.
3. The use of directivity in the attribute extraction
4. The use of supervised and unsupervised neural
networks to recombine the attributes into new
attributes with improved separation power.
5. The possibility to iterate the process by first enhancing
the texture of the objects then detecting them by either
manual interpretation, or automated detection after
application of edge detection and pattern recognition
The chimney cube: Part I; methodology
6. The ability to extract the detected objects and zoom in
for detailed characterization work.
A specific application of the method was presented: the
chimney cube. This cube may add a new dimension to
seismic interpretation as an indirect hydrocarbon detector.
The interpretation of these cubes is the subject of part 2 of
Den Norske Stats Oljeselskap a.s (Statoil) is gratefully
acknowledged for the use of their data and the permission
to publish this paper.
Meldahl, P., Heggland, R., de Groot, P.F.M. and Bril, A.H.,
1998, Method of Seismic Body Recognition. Patent
application GB 9819910.02
Steegs, T,P,H., 1997. Local Power Spectra and Seismic
Interpretation. Ph.D thesis Section for Applied Geophysics,
faculty of Applied Earth Sciences, Delft University of
Technology. ISBN 90-90108 12-2
Heggland, R., 1997. Detection of Gas Migration from a
Deep Source by the Use of Exploration 3D Seismic data,
Marine Geology, 137, 41-47.
Heggland, R., 1998. Gas Seepage as an Indicator of Deeper
Prospective Reservoirs. A Study on Exploration 3D
seismic. Marine and Petroleum Geology, 15, 1-9.
Heggland, R., 1999. The chimney cube, an example of
semi-automated detection of seismic objects by directive
attributes and neural networks: Part II; interpretation
de Groot, P.F.M., 1999a. Seismic Reservoir
Characterisation Using Artificial Neural Networks. 19th
Mintrop-Seminar, 16 – 18 May 1999, Münster, Germany.
de Groot, P.F.M., 1999b. Volume transformation by way of
neural network mapping. 61st EAGE Conference, Helsinki,
7-11 June 1999.
Meldahl,P., 1998. Survey Evaluation and Design,
prediction of resolution versus line interval.
The Leading Edge, November 1998.
Fig. 3 Seismic line with seed interpretation showing
locations inside a chimney and outside the chimney.
Fig. 4 Neural network topology. Attributes are
extracted in 3 vertically aligned windows relative to the
reference time (which in itself is also used as an input
variable to the network).
Fig. 5 One inline through the output chimney cube
(compare with Fig. 3).