244
10: Pakorn Watanachaturaporn, Manoj K. Arora
Multiclass SVM using Linear Kernel
100
90
80
>u
70
f!
::J
60
u
u
50
4(
~
40
II
30
>
0
20
10
DO
~
~
~
...
/
T /"
~
/ r r
""E>
/,/
/
V
Ii
~Pairwise
IK'
"
-B-OAG
. ~
!
~ Qne.Y·Rest
J
~ Qne.Y·RestBaI
-"~.~. ~
.~.
Penalty Value (C)
Fig.lO.2. Overall accuracy of multi class approaches applied on the multispectral image
Multiclass SVM using Linear Kernel
100
90
80
>.
u
70
E
::J
60
u
u
50
4(
e 40
. . 30
>
0
20
10
0
#" #''' ~"
<:)"
" "
,,<:)
~ ,,#' ,,# ,,/
<:) .
<:).
<:).
<:) .
<:).
Penalty Value (C)
Fig.lO.3. Overall accuracy of multiclass approaches applied on the hyperspectral image
It can be seen that for both the datasets, the pairwise and DAG methods result in significantly higher accuracies than the other two methods. Both these
methods show a similar trend with accuracy reaching 95%, even for a small
value of C. The other two methods could attain an accuracy of only 70%. In
all the methods, as the value of C is increased, the accuracy increases but gets
saturated at a certain value of C in case of multispectral data. However, in the
hyperspectral dataset, the accuracy drops after attaining a maximum value for
each method. This shows that there is an optimum value of C irrespective of