Support Vector Machines for Classificationof Multi- and Hyperspectral Data
100
90
80
>- 70
" f! ::I 60
" " 50
E 40
CD
> 30
0
20
10
o
Sigmoid Kernel
/~
?
\
r
\
/
~
/
.~.
/b
"().
~
"Penalty Value (C)
v
251
Fig.1O.14. SVM performance using the sigmoid kernel applied on the multispectral image
Linear Kernel
100
90
80
>"
70
E
::I
60
" "
50
E
40
CD
30
>
0
20
10
/'
~
?
~
/
\
~
\
Af'
~
~
\ \
~
~
Penalty Value (C)
Fig.IO.IS. SVM performance using the linear kernel applied on the hyperspectral image
example, for the polynomial kernel with degree 5, an accuracy of 92% is
achieved with a C value as low as 10- 5 • This shows that if the degree of the
polynomial is high, a lower C value can be adopted. However, polynomials
with lower degrees 0. e. 2 to 4) achieved higher accuracy than those obtained
from polynomials with higher degrees (i. e. 5 to 7). Hence, a trade-off between
the C value and the degree of polynomial may be maintained to get maximum
classification accuracy for hyperspectral data. The linear kernel, although it
attains a maximum accuracy of95% at a certain C value, it drops dramatically