information drive an equivalent form, called Minimal Redun-
dancy Maximal Relevance (
MRMR
) criterion. The
FS
approach
proposed by Meyer and Bontempi (2006) called “Double
Input Symmetrical Relevance (DISR)” is based on information
theoretic and uses variable complementary. They have exam-
ined their proposed approach for very large number of feature
vectors. Fleuret (2004) has proposed a high speed scheme in
the context of classification called “Conditional Mutual In-
formation Maximization (
CMIM
)” that does not select a feature
similar to already picked ones, and it does not carry addition-
al information about the class to predict. Thus, this criterion
ensures a tradeoff between independency and discrimination.
Battiti (1994) has proposed “Mutual Information Feature
Selection (
MIFS
)” that evaluates
MI
between individual features
and class labels, instead of calculating the joint
MI
between the
selected feature set and the class variables. It selects the features
that have maximum
MI
with the class labels and less quantity
compared to the accumulated
MI
with the previously selected
features. Cheng
et al
. (2011) have proposed “Conditional
Mutual Information Feature Selection (
CMIFS
)” based on the
link between interaction information and conditional mutual
information. It takes into account both redundancy and synergy
interactions of the features. Their evaluation results show that
CMIFS
achieves better classification accuracy compared to
MIFS
,
in same condition. An approach was proposed by Yang and
Moody (1999) using the “Joint Mutual Information (
JMI
)” that
focuses on increasing complementary information between fea-
tures.
JMI
provides trade off in terms of accuracy, stability, and
flexibility with small data samples (Brown
et al.
, 2012). The ap-
proach “Interaction Capping (
ICAP
)” proposed by Jakulin (2005)
uses the min/max strategy instead of a linear combination of
features. Lin and Tang (2006) have proposed “Conditional In-
fomax Feature Extraction (
CIFE
)” that drives a new information
decomposition model in which both relevancy and redundancy
are considered. It maximizes the joint class-relevant information
by reducing the class-relevant redundancies among the features.
In contrast to above methods, some methods use statisti-
cal approach and avoid heuristic search. For example, Relief
method proposed by Kira and Rendell (1992) does not con-
sider redundant features and is not affected by feature interac-
tion. If most of the given features are relevant to the concept,
it would select all of them, even though if only a small num-
ber of them are sufficient for the concept description.
Most of
FS
methods are dependent to the dataset. They are
based on search methods for selecting appropriate feature sets
that is very time consuming. Moreover, they usually gener-
ate all features blindly that may result in redundant features
which should be reduced in a feature reduction step. In this
paper, a
FS
method for segmentation of
VHRSI
from agricul-
tural Area of Interest (
AOI
) that is not dependent to a specific
satellite is proposed. In this regard, the effective feature types
based on reasonable facts of agricultural areas is defined and
appropriate candidate features for each feature type without
any search are selected. In the proposed method, the features
are selected in a supervised fashion by a sample image as
a representative of similar regions on the Earth. Then, the
selected model including feature types along with clustering
method can be used for segmentation of images captured by
different satellites with similar spatial resolution in specified
regions (e.g., agricultural areas in this paper).
The proposed
FS
method includes three steps: (a) feature
type selection, (b) candidate selection, and (c) feature set
selection. The output of the Step c is a selected model which
includes a pair of a feature set and a clustering method. These
three steps procedure is applied to only one sample image
as a labeled data. By using the selected model (a feature set
and a clustering method), other VHRSIs with same land cover
classes from satellites with similar spatial resolution can be
segmented. The performance of the proposed method with
aforementioned
FS
methods is compared. The comparison of
the results of image clustering based on these methods and
the proposed method shows that the proposed
FS
method
leads to better results.
The remainder of this paper is organized as following. In
the next section, the proposed
FS
method is presented and
explained in detail, followed by the performance of the pro-
posed method and evaluation through experimental results.
The final section is our conclusions..
The Proposed FS Method
An ideal model for image segmentation based on clustering
approach includes two steps; predetermine efficient fea-
ture types and a suitable clustering method. To achieve this
model, a framework is proposed as shown in Figure 1.
As shown in Figure 1, the output of the model selection
block is a clustering model for a given labeled image. The
selected clustering model includes a suitable feature set and
a clustering algorithm. By using the selected model, other
VHRSIs with same land cover classes from satellite data with
almost same spatial resolution can be segmented. In the pro-
posed method, the
FS
model is applied only on a training data
(called here after labeled image) and selected model (a feature
set and a clustering algorithm) are used for segmentation of test
datasets. For model selection, the essence and reasonable fea-
ture types is selected in the first step. Some feature types have
multiple directions. Therefore, the suitable direction is se-
lected for these feature types in the second step. For example,
“Energy” is a feature type that is calculated from any direction
of
GLCM
, so one feature from these four candidates should be
selected. In the third step, clustering of the feature sets is done
with three clustering methods including K-means,
FCM
, and
ISODATA
. It is worthy to note that the efficiency of the selected
features is investigated by well-known clustering methods;
however, the segmentation results can be improved by other
clustering methods. As mentioned in the Introduction, these
three clustering methods commonly are used by researchers.
The results of the clustering methods is evaluated and
compared by a quantitative metric. Based on this comparison,
appropriate clustering model (a feature set and a clustering
algorithm) is selected to be used for segmentation of the other
images. The selected model can be used for segmentation of
images with similar land cover to that of the labeled image.
The proposed
FS
is concluded on panchromatic labeled im-
age. The selected features are generated using other bands to
yield the feature vector. The Steps of Figure 1 are addressed
in details as follows.
Figure 1. Flowchart of VHRSI segmentation framework for implementation of the proposed FS method.
214
March 2016
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
