IEEE Signal Processing - March 2018 - 90

gas-chimney detection in two ways. First, it limits the amount of
facies analysis has gained popularity, particularly for a deposition
training data, and thereby the performance of supervised learnsystem with complexities, and automatically classified all subtle
ing may be affected. Weakly supervised and unsupervised methstructures, including the meandering channel, overbanks, fans,
ods would be more applicable. Second, it increases the sensitivity
and lobs. The use of computational classification began soon after
to seismic noise. Reflection pattern-based learning could help
the development of seismic attributes in the 1970s with the work
improve the noise robustness of gas-chimby Justice et al. [33]. Barnes and Laughlin
ney detection, such as the chaotic labeling
[34] reviewed several unsupervised classifiThe focus of this
described in the next section given the chaotic
cation techniques, including k-means, fuzzy
article is to present the
reflections in a gas chimney.
clustering, and self-organizing maps (SOMs)
seismic interpretation
and emphasized the importance of seismic attributes over the classifiers. Wallet et al. [35]
Subsurface labeling and classification
in the framework of the
developed the generative topographic mapIn the previous section, we provided a review
generalized pipeline for
ping for unsupervised waveform classificaof seismic interpretation for subsurface event
imaging data analysis,
tion. Song et al. [36] combined multilinear
detection and tracking. Although the techwhere both human visual- niques have significantly reduced the time
subspace learning with the SOMs for imsystem- and learningproved seismic facies analysis in the presence
and effort required for manual interpretation,
based models have the
of noise. A comprehensive study of both suthere is at least one aspect involved that is still
pervised and unsupervised facies analysis can
done manually, namely, the manual process
potential to assist in
be found in [37]. With these depositional feaof extracting subvolumes from a given data
discovering underlying
tures in a channel system well differentiated
volume based on their dominant subsurface
structures that may not
by either seismic attributes or facies analysis,
structure, so that detection or tracking can
have been discovered
they could then easily be extracted as sepabe performed on the extracted data. In this
otherwise.
rate geobodies by seeded tracking. However,
section, we discuss a framework we recently
reliable differentiation of various depositional
developed to address this issue. With this
features (e.g., overbank, delta, and levee) remains challenging,
framework, we attempt to eliminate the aforementioned bottledue to the proximity and overlying distribution in space between
neck and streamline the interpretation process by building on the
each other and, more importantly, their similar reflection patterns
recent advances in semantic segmentation and scene labeling.
in 3-D seismic data. Such a goal can be achieved by analyzing the
seismic images at a smaller scale to capture the subtle differences
General framework
between various features.
Seismic volume labeling is the process of classifying each
voxel in a given seismic volume into one of many predefined
structures. This process can help classify entire seismic volGas-chimney detection
umes into regions of interest that contain specific subsurface
In seismic sections, a gas chimney is visible as vertical zones of
structures. These regions can then be extracted, and various
poor data quality, chaotic reflections, or push-downs. Therefore,
interpretation algorithms can be applied to these regions for
it can be detected using attributes similar to those used for saltmore refined results.
dome detection, such as the coherence [28] and the GoT [24]. The
While a variety of influential works have been proposed in
major difference of gas chimneys is the sparse distribution in a
the area of semantic segmentation [43]-[47], seismic data presseismic volume, and thereby manual identification and interpreents challenges that cannot be immediately solved by the existtation of them is labor intensive, and the conventional tracking
ing methods. First, unlike natural images, where edges between
tools described in the section "Fault and Salt-Dome Tracking"
objects are well defined, edges between subsurface structures
may also fail in detecting the chimneys that are isolated from each
in seismic data are either not well defined or are characterized
other. However, the machine-learning-based approach offers an
by a change in overall texture rather than a sudden change in
efficient solution to such limitation. For example, Heggland et
amplitude. Second, unlike natural images, seismic data is gray
al. [38] combined a set of seismic attributes and the multilayer
scale, and, thus, color features cannot be used to distinguish
perception (MLP) to create a chimney cube for a semiautomatic
various structures. Furthermore, there is a severe lack of both
detection, which has been applied to multiple data sets, such as
labeled seismic data for training and well-established benchthe F3 block [39] and the Taranaki basin in New Zealand [40]. In
marks for testing various learning-based approaches. This is
recent years, researchers also have tried more advanced machinepartly due to the intellectual property concerns in the oil and
learning algorithms. For example, Xiong et al. [41] applied adapgas industry. Also, because of the lack of ground truth and the
tive boosting (AdaBoost) to the design of the optimal learning
subjective nature of manual interpretation, it is often difficult
algorithm for identifying gas chimneys, which generated more
for different geophysicists to agree on the same interpretation
reliable results than the k-nearest neighbor method. Xu et al. [42]
for a given volume.
implemented the sparse autoencoder for gas-chimney detection,
Naturally, machine-learning techniques are well suited
and the accuracy is greatly improved compared to the traditional
approaches for seismic volume labeling. However, the lack of
MLP algorithm. The sparsity of the spatial distribution of gas
labeled training data poses a significant challenge. To tackle this
chimneys in a seismic data set adds the difficulty of reliable
90

IEEE Signal Processing Magazine

March 2018

Table of Contents for the Digital Edition of IEEE Signal Processing - March 2018