3.3 Results
Outputs of the GA-SVM model include optimized kernel
parameters and feature subset, classification result, evaluation
of the classification accuracy, as shown in Table 1. The
classification result is shown in Figure 4. The evolution of
fitness value with number of generation is shown in Figure 5.
It can be seen in Table 1 that the number of features, i.e. bands,
selected for the use of classification was only 13, much smaller
than the total number of bands (198). Classification accuracy
was 92.51% when using the optimized kernel parameters and
feature subset, comparing to 88.81% when no feature selection
was performed. The optimized values of the two SVM kernel
parameters were 95.0297 and 0.2021, respectively, different
from the default values of 100 and 0.005 in ENVI, which again
proved the necessity of optimizing the kernel parameters.
As can be seen in Fig. 5, the fitness value increases with
number of generation. There are sharp increase at the 3rd and
30th generation, although the fitness value may remain
unchanged during several generations. The acceptable fitness
level was reached at the 30th generation.
Although the maximum number of generation evolved was only
40 in our study due to limited computer memory. The GA-SVM
method has proved to be able to return acceptable result under
the limited memory resource. It may be a safe assumption that
provided with larger computer memory, better results may be
gained.
73. 20
In conclusion, the proposed wrapper feature selection method
GA-SVM can optimize feature subsets and SVM kernel
parameters at the same time, therefore can be applied in feature
selection of the hyper spectral data.
Bengio Y., 2000. Gradient-based Optimization of Hyper
parameters. Neural Computation, 12(8), pp. 1889-1900.
Bi J., Bennett K.P., Embrechts M., Breneman C.M., Song M.,
2003. Dimensionality reduction via sparse support vector
machines. J. Mach. Learn. Res. 3, pp. 1229-1243.
Bradley P. S., Mangasarian O. L., Street W. N., 1998. Feature
selection via mathematical programming. INFORMS Journal
on Computing, 10, pp. 209-217.
Chen M., Yi Y. H., Liu Zh. G., Li D. R., QIN Q. Q., 2006.
Study of Band Selection Methods of Hyperspectral Image
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39
Number of Generation
Figure 5. Change of fitness with population evolvement