16 training samples, and 8 extracted features are shown in
Table 3. The classification results obtained by the
SVM
clas-
sifier using 32 training samples and 7 extracted features are
shown in Table 4. We repeat these experiments for the Pavia
dataset, and the experimental results with 16 training samples
are shown in Tables 5 and 6. An intuitive view for the Pavia
image and the classification maps obtained by the
SVM
clas-
sifier, 16 training samples and 6 features are shown in Figure
5. The average classification accuracy versus the number of
extracted features for (a) the
KSC
dataset using the
SVM
classifi-
er, (b) the
KSC
dataset using the
ML
classifier, (c) the Botswana
dataset using the
SVM
classifier, and (d) the Botswana dataset
using the
ML
classifier are shown in Figure 6. The results of
MCNemars test for all discussed cases are shown in Table 7.
The comparison of
2DLDA
with
LDA
in different size of training
samples is done with a fixed number (=8) of extracted features
and by the (a)
SVM
, and (b)
ML
classifiers for the Indian data-
set; the obtained results are shown in Figure 7.
T
able
5. T
he
R
esults
of
C
lassification
A
ccuracy
O
btained
by
the
SVM C
lassifier
, 16 T
raining
S
amples
,
and
S
ix
F
eatures
for
the
P
avia
D
ataset
class
2DLDA
LDA
NWFE
GDA
No. of classes
Name of class
# samples Acc.
Rel.
Acc.
Rel.
Acc.
Rel.
Acc.
Rel.
1
Asphalt
6631
0.73
0.93
0.48
0.29
0.65
0.75
0.67
0.87
2
Meadows
18649
0.54
0.84
0.55
0.82
0.57
0.86
0.64
0.82
3
Gravel
2099
0.79
0.66
0.33
0.20
0.58
0.51
0.75
0.31
4
Trees
3064
0.94
0.44
0.72
0.62
0.95
0.47
0.95
0.44
5
Painted metal sheets
1345
0.99
0.97
0.83
1.00
0.98
0.98
0.98
0.96
6
Bare Soil
5029
0.60
0.38
0.24
0.26
0.64
0.39
0.49
0.44
7
Bitumen
1330
0.90
0.51
0.38
0.18
0.84
0.40
0.89
0.44
8
Self-Blocking Bricks
3682
0.81
0.78
0.20
0.30
0.67
0.85
0.16
0.62
9
Shadows
947
1.00
0.99
0.82
0.58
1.00
0.99
1.00
1.00
Average Acc. and Average Rel.
0.81
0.72
0.50
0.47
0.76
0.69
0.73
0.66
Overall Acc.
0.93
0.89
0.93
0.93
Kappa coefficient
0.60
0.35
0.57
0.55
T
able
6. T
he
R
esults
of
C
lassification
A
ccuracy
O
btained
by
the
ML C
lassifier
, 16 T
raining
S
amples
,
and
S
ix
F
eatures
for
the
P
avia
D
ataset
class
2DLDA
LDA
NWFE
GDA
No. of classes
Name of class
# samples Acc.
Rel.
Acc.
Rel.
Acc.
Rel.
Acc.
Rel.
1
Asphalt
6631
0.79
0.92
0.21
0.36
0.70
0.82
0.71
0.90
2
Meadows
18649
0.62
0.85
0.48
0.80
0.67
0.86
0.62
0.88
3
Gravel
2099
0.83
0.59
0.18
0.13
0.55
0.37
0.72
0.51
4
Trees
3064
0.98
0.48
0.80
0.56
0.97
0.47
0.98
0.47
5
Painted metal sheets
1345
0.99
0.98
0.97
0.98
1.00
0.99
1.00
0.98
6
Bare Soil
5029
0.60
0.41
0.26
0.21
0.59
0.48
0.66
0.45
7
Bitumen
1330
0.76
0.70
0.39
0.11
0.78
0.61
0.77
0.65
8
Self-Blocking Bricks
3682
0.71
0.76
0.38
0.22
0.59
0.71
0.74
0.73
9
Shadows
947
0.96
1.00
0.78
0.40
0.98
1.00
0.98
1.00
Average Acc. and Average Rel.
0.80
0.74
0.49
0.42
0.76
0.70
0.80
0.73
Overall Acc.
0.94
0.88
0.94
0.94
Kappa coefficient
0.64
0.31
0.61
0.63
T
able
7. T
he
R
esults
of
MCN
emars
T
est
(E
ach
C
ase
of
the
T
able
R
epresents
Z
rc
where
r
is
the
R
ow
and
c
is
the
C
olumn
)
Indian, SVM classifier, 16 training samples, 9 extracted features
2DLDA LDA NWFE
GDA
2DLDA
0
54.95
15.91
20.11
LDA
-54.95
0
-44.86
-42.22
NWFE
-15.91
44.86
0
4.08
GDA -20.11
42.22
-4.08
0
Indian, ML classifier, 16 training samples, 8 extracted features
2DLDA LDA NWFE
GDA
2DLDA
0
59.77
10.02
11.04
LDA
-59.77
0
-53.99
-53.73
NWFE
-10.02
53.99
0
0.88
GDA -11.04
53.73
-0.88
0
Indian, SVM classifier, 32 training samples, 7 extracted features
2DLDA LDA NWFE
GDA
2DLDA
0
26.73
14.37
34.92
LDA
-26.73
0
-14.21
6.52
NWFE
-14.37
14.21
0
21.91
GDA -34.92
-6.52
-21.91
0
Indian, ML classifier, 32 training samples, 8 extracted features
2DLDA LDA NWFE
GDA
2DLDA
0
15.89
7.76
4.53
LDA
-15.89
0
-9.44
-12.30
NWFE
-7.76
9.44
0
-3.70
GDA
-4.53
12.30
3.70
0
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
October 2015
783
