A Feature Selection Approach for Segmentation
of Very High-Resolution Satellite Images
Ahmad Izadipour, Behzad Akbari, and Barat Mojaradi
Abstract
Most of the feature selection (
FS
) methods in the literature
determine features that are appropriate only for a given dataset.
In contrast, in this paper a
FS
method that is not dependent to a
specific dataset is proposed. In this regard, the effective feature
types based on reasonable facts are predefined and appropriate
candidate features for each feature type are selected. In pro-
posed method, the features selected from a single labeled image
can be used in segmentation of images captured by different sat-
ellites with similar spatial resolution. The selected feature types
contain spatial and spectral features. The selected features are
applied for segmentation of the images captured by QuickBird
and GeoEye satellites and obtained results of proposed method
are compared with well-known
FS
methods. Using different
evaluation measures, our comparison shows the efficiency of
the proposed method in providing better segmentation com-
pared to other
FS
methods that are presented in this paper.
Introduction
The recent availability of new satellite sensors that are capable
of providing Very High Resolution Satellite Images (
VHRSI
) has
concentrated the experts to extract the information from
VHRSI
efficiently. In this way, image segmentation is fundamental
work. Therefore, it is necessary to develop an efficient image
segmentation approach to extract accurate segments from
VHRSI
. The segmentation task that partitions the images into
non-overlapping homogenous regions is a low level image pro-
cessing technique for many applications in the remote sensing
domain such as object detection, change detection, topograph-
ic cartography, and cartography of land occupation (Congalton,
2010). The complexity and redundancy are challenging in seg-
mentation of
VHRSI
due to more spatial resolution and detailed
information about spectral range, shape, context, and texture of
the images. There are many kinds of segmentation approaches
in the literature that have been developed for satellite images,
such as graph-based methods (e.g., Tilton
et al
., 2008), region
growing methods (e.g., Baatz and Schape, 2000; Tilton, 1998;
Gao
et al
., 2011), evolution methods (e.g., Bhandarkar and
Zhang, 1999; Das
et al
., 2006) and clustering methods (e.g., Liz-
arazo and Barros, 2010; Bitam and Ameur, 2013).
Among the above segmentation methods, clustering-based
methods are more suitable when more spectral and texture
features are available. Feature generation plays a main role in
clustering algorithms and depends on the application used
domain. The result of any clustering method highly depends
on the number and types of generated features, and there-
fore the clustering methods using different feature types and
feature numbers achieve different results. One of the most
important feature generation schemes is based on Gray Level
Co-occurrence Matrix (
GLCM
) for texture image segmentation
(e.g., Xiangrong Zhang
et al.
, 2008).
K-means and fuzzy C-means (
FCM
) clustering algorithms
have been used for image segmentation in a number of re-
searches (e.g., Rekik
et al.
, 2006; Rekik
et al.
, 2009; Cannon
et
al.
, 1986; Gen-yuan
et al.
, 2009; Bitam and Ameur, 2013). The
main problem of both methods is the number of initial clusters
that must be manually determined by the user. In contrast,
ISO-
DATA
clustering is an unsupervised clustering that determines
the number of clusters dynamically.
ISODATA
clustering is simi-
lar to K-means clustering, but the
ISODATA
algorithm uses some
constrains on splitting and merging of clusters to obtain the
optimal partition. In
ISODATA
, clusters are merged if the num-
ber of members in a cluster is less than a predefined threshold
or the centers of two clusters are closer than a predefined
threshold. If standard deviation of the cluster exceeds a pre-
defined value or the number of members is more than a thresh-
old then the cluster is split into two clusters. For example,
Cai-hong
et al
. (2012) determined the initial cluster number
by “Particle Swarm Optimization (
PSO
)” for
ISODATA
clustering
approach. These three clustering methods have been used in
many works in the literature (e.g., Awed
et al.
, 2007; Hung
et
al
., 2008; Cai-hong
et al
., 2012) as a base clustering method or
for comparing with the proposed methods. We also have used
these clustering methods in the proposed method.
Since, high-dimensional feature vectors is one of the main
challenges of clustering methods, feature reduction is an
essential step in image clustering especially for
VHRSI
. Some
works (e.g., Akçay and Aksoy, 2008) have employed a com-
bination of all features (e.g., Principal Component Analysis
(
PCA
)) in which each component of the reduced vector is a
function of all features of the original feature vector. Some
of other feature reduction schemes (e.g., Mitra
et al.
, 2002,
Moustakidis
et al.
, 2012 and Xie
et al.
, 2013) select effective
features from original feature vector. Feature Selection (
FS
)
techniques can be generally categorized into two approaches
that are classifier-dependent (“wrapper” and “embedded”
methods), and classifier-independent (“filter” methods).
Wrapper methods utilize the machine learning approach to
select feature sets. Filter methods select subset of features
based on a predefined measure and remove low scoring
features. In spite of the better performance, wrapper methods
can be computationally expensive and have a risk of over-fit-
ting to their learning model (Brown
et al.
, 2012).
In the filter methods for
FS
, the evaluating measures take
a critical role in detecting feature relevancy and redundancy.
“Mutual Information (
MI
) (Shannon, 1948)” is a most import-
ant measure that has been used in the most of the
FS
methods
in the literature (e.g., Lewis, 1992; Peng
et al.
, 2005; Meyer
and Bontempi, 2006). These methods uses heuristic search to
select appropriate feature sets. Peng
et al
. (2005) to reach the
maximal statistical dependency criterion based on mutual
Ahmad Izadipour and Behzad Akbari are with the Faculty of
Electrical and Computer Engineering, Tarbiat Modares Uni-
versity, Tehran, Iran (
).
Barat Mojaradi is with the School of Civil Engineering, Iran
University of Science and Technology, Tehran.
Photogrammetric Engineering & Remote Sensing
Vol. 82, No. 3, March 2016, pp. 213–222.
0099-1112/16/213–222
© 2016 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.82.3.213
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
March 2016
213
