IEEE Signal Processing - March 2018 - 91

years for image analysis. They are also potential candidates for
problem, we have generated the LANDMASS [48] data set that
the labeling task.
contains more than 17,000 seismic images extracted from The
When applied to image processing problems, it is usually
Netherlands North Sea F3 block [49]. The images are grouped
desirable for a texture attribute to possess properties such as
into four classes based on the subsurface structure they
illumination-, rotation-, and scale-invariance for better robustcontain: horizons, chaotic horizons, faults, and salt
ness. However, with seismic data, this is not
domes. Although the data set contains these
always the case. For example, a vertical slice
specific structures, the framework for seismic
The ultimate goal of
of a seismic volume (or a seismic section) is
labeling discussed next can be extended to
seismic volume labeling
characterized by strong directionality, with
other seismic structures as well.
is to efficiently and
horizons typically extending in the horizontal
In [50], Alaudah and AlRegib proposed
accurately classify entire
direction and faults in the vertical direction.
using reference exemplars and seismic imIn such cases, being rotation-invariant is no
age retrieval to label seismic volumes in
seismic volumes based
longer a critical requirement for the attributes.
a weakly supervised fashion. First, given
on their subsurface
Another important difference between a seisa few hand-selected exemplar images,
structures such as those
mic image and a natural texture image is that
X = [x 1, x 2, f, x N e], that contain subsurface
shown in Figure 8.
some subsurface structures (e.g., faults) are
structures belonging to the different classes of
u , is generof very fine-scale along certain dimensions,
interest, an augmented data set, X
which is not typical with natural textures. Thus, it is important for
ated in an unsupervised fashion using similarity-based retrieval.
a texture attribute to be able to capture such fine details.
This is done to obtain enough data to train a supervised machineRecently, some comparative studies were conducted to examlearning model. Then, various features or attributes are extracted
ine various texture attributes in the context of seismic volume
from these images to train a classifier. The section "Oversegmenlabeling. In one study, the focus was on a group of spatial attritation of Seismic Volumes" describes seismic volume segmentabutes from the family of local descriptors, including the local
tion and its use to enforce the local spatial correlation of the labels
binary pattern, a few of its typical variants, and the local radiand improve the computational efficiency of the approach. Fius index [51]. These attributes can represent texture patterns
nally, the sections "Labeling of Seismic Volumes" and "Weakly
with robustness and computational efficiency. For comparison
Supervised Pixel-Level Annotation" describe various methods
purposes, the study also included two traditional seismic attrito obtain the final labeling of the seismic volume. The overall
butes in the spatial domain, i.e., the GLCM and the semblance.
framework is outlined in Figure 7.
According to the study, the local descriptors and the GLCM are
all good attributes for labeling seismic volumes. However, each
Building blocks
attribute displayed different characterizing capabilities for differIn the remaining part of this section, we describe the major coment subsurface structures. Thus, they should be selected accordponents of this weakly supervised approach for seismic volingly if there is a preference for certain structures to be labeled
ume labeling.
with more reliability.
In a separate study [52], multiresolution attributes in the freTexture attributes for seismic labeling
quency
domain were examined for seismic volume labeling,
Seismic images are often well characterized by texture features,
including the discrete wavelet transform and its nonsubsampled
or texture attributes, mainly because they are textural in nature.
version, Gabor filters, the steerable pyramid, the contourA few classical texture attributes were explored for traditional
let transform and its nonsubsampled counterpart, and the curvetasks such as salt-dome detection, but they need to be further
let transform. Effective singular values are extracted from each
examined in the context of seismic labeling. In addition, there are
transformed subband and then concatenated into a feature vector.
a great number of advanced texture features developed in recent

Training Stage

Hand-Labeled
Exemplars

3-D Seismic
Volume

Similarity-Based
Retrieval

Feature
Extraction

Classifier

Oversegmentation

Feature
Extraction

Classifier
Labeled 3-D
Seismic Volume

Labeling Stage

Figure 7. A block diagram illustrating the weakly supervised seismic volume-labeling approach described in the section "Subsurface Labeling and
Classification."
IEEE Signal Processing Magazine

March 2018

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