Support Vector Machines for Classificationof Multi- and Hyperspectral Data
249
Linear Kernel
100
90
eo
~ 70
I!!
:I
60
u
~ 50
~
40
~
30
20
~~
/
I
I
I
1
/
~
10
Penalty Value (C)
Fig.lO.lO. SVM performance using the linear kernel applied on the multispectral image
Polynomial Kernel of degrees 2 to 4
100
90
= ......
~ 12rW""'<7
eo
/rjf /
~ 70
E 60
/ / ?
:I
/ t2f 1
!:l 50
C(
/ / /
~
40
'/ I
~PoIy02
&
30 J Jb
-B-PoIy03
20 I~
-&-PoIy04
10
0
#" #" '
I)"
I)'
, "I) ,, ,#' ,#' #'
I)' "
Penalty Value (e)
Fig.IO.II. SVM performance using the polynomial kernels of degrees 2 to 4 applied on the
multispectral image
there is a gradual increase in accuracy for small increases in C values. This
demonstrates that if a polynomial with a higher degree is used as the kernel, higher accuracy can be achieved at very low C values. In fact for polynomials with higher degrees, the accuracy may decrease after attaining the
maximum at a certain C value unlike the polynomials with lower degrees,
where the accuracy gets saturated. In case of the RBF kernel also, the accuracy
increases as the penalty value increases but drops marginally after attaining