condition by using the “FEAST” toolbox (Brown
et al.
2012).
As mentioned in the Introduction, the compared methods
need three parameters for
FS
including: (a) the original feature
vector, (b) labeled data (here it corresponds to the labeled
image), and (c) the number of final features (except
FCBF
and
Relief methods that ignore the number of final features).
Note that the
FCBF
method searches the optimum number of
features and the Relief method ranks all features in a descend-
ing sort. Therefore, these methods are applied on DS1 which
is a labeled image and the optimum features, except
FCBF
,
to be selected are set to the feature numbers in the proposed
method. The achieved results are given in Table 7. Accord-
ing to the obtained results, the proposed
FS
method provides
maximum
β
/
F
compared to other
FS
methods. This dem-
onstrates that in the same condition, the proposed method
yielded better results. Our proposed method can not only
achieve better results on the labeled images, but it also can be
applied on other datasets. To show this, the next experiments
have been conducted.
To evaluate the efficiency of the selected features for general-
ization and utilizing them for other similar datasets, the differ-
ence of spatial resolution should be taken into consideration. In
this regard, the mask size may play main role to resolve spatial
resolution difference. Therefore the selected features from DS1
(labeled image with spatial resolution of 65 cm) are used to
cluster DS7 (with spatial resolution of 41 cm) with a different
W. According to the obtained results that are given in Table 8,
the larger W yields larger
β
. In contrast,
F
index becomes almost
maximum when W is 5 which demonstrates the maximum
inter-cluster distance and the maximum over-segmentation.
These effects almost apply to other datasets. On the other hand,
the
β
/
F
value is maximum when W is 3 for all methods. More-
over, Figure 8 demonstrates the effect of W on clustered images.
Since with increasing W the edges of land cover are destroyed
T
able
6. V
arious
FS M
ethods
are
U
sed
for
C
omparison
Row Abbreviation
Full name of
the FS method
Author(s)
1 CMIM
Conditional Mutual Info Maximization
Fleuret (2004)
2 MRMR
Max-Relevance Min-Redundancy
Peng et al.(2005)
3 ICAP
Interaction Capping
Jakulin (2005)
4 DISR
Double Input Symmetrical Relevance
Meyer and Bontempi (2006)
5 CIFE
Conditional Infomax Feature Extraction
Lin and Tang (2006)
6 CMIFS
Conditional MIFS
Cheng et al.(2011)
7 MIFS
Mutual Information Feature Selection
Battiti (1994)
8 JMI
Joint Mutual Information
Yang and Moody (1999)
9 FCBF
Fast Correlation Based Filter
Yu and Liu (2004)
10 MIM
Mutual Information Maximization
Lewis (1992)
11 Relief
Relief
Kira and Rendell (1992)
T
able
7. C
omparison
of
P
roposed
M
ethod with
other
M
ethods
on
DS1 D
ata
(L
abeled
I
mage
)
Index Proposed CMIM MRMR ICAP DISR CIFE CMIFS MIFS JMI
FCBF MIM Relief
β
8.098
7.06
6.73
6.85
9.27
6.8
7.02
8.05
8.05
7.34
8.0
10.8
F (10
7
)
1.27
4.72
5.82
5.42
7.58
5.72
4.57
7.15
7.16
7.28
7.16
5.92
β
/ F (10
-7
)
6.37
1.21
1.42
1.26
1.22
1.20
1.53
1.12
1.12
1.00
1.12
1.83
T
able
8. C
omparison
of
the
P
roposed
M
ethod with
O
ther
M
ethods
on
DS7 (G
eo
E
ye
D
ataset
)
Proposed CMIM MRMR ICAP DISR CIFE CMIFS MIFS JMI
FCBF MIM Relief
β
W=3 7.2245 5.83
5.7
5.97
7.2
6.01
5.83
6.47
6.57
6.63
6.47
8.97
W=5
9.64
7.69
7.62
7.74 10.01 7.92
7.63
8.96
8.96
8.48
8.96 11.18
W=7 10.91 8.50
8.34
8.40 11.54 8.57
8.40 10.15 10.15 8.74 10.15 11.97
F
(10
8
)
W=3
1.18
1.70
3.26
2.37
2.53
3.06
1.70
2.05
2.46
7.16
2.02
1.27
W=5
9.4
9.58 11.87 11.90 9.78 12.92
8.2
10.52 10.52 11.13 10.53 11.32
W=7
2.95
5.74
7.54
7.55
4.74
7.68
6.54
5.66
5.64
7.17
5.66
3.31
β
/F
(10
-8
)
W=3
6.12
3.43
1.75
2.52
2.84
1.96
3.43
3.15
2.67
0.93
3.2
7.06
W=5
1.02
0.80
0.64
0.65
1.02 0.613 0.93
0.85
0.85
0.76
0.85
0.98
W=7
3.69
1.48
1.10
1.11
2.43
1.11
1.28
1.79
1.79
1.21
1.79
3.61
(a)
(b)
(c)
Figure 8. Results of K-means clustering algorithm on DS5 with multiple masks: (a) W = 3, (b) W = 5, and (c) W = 7.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
March 2016
219
