Automated classification provides an efficient and accurate way to map land cover
information classes especially highly dispersed covers such as vegetation. Researchers in remote
sensing keep improving classifiers for land cover and land use classification. Many parametric
schemes [28], where decision boundaries are found after distribution functions are estimated from
given training sets, have been presented. However, the classification results are often not
satisfactory since the estimated distribution function, which is usually Gaussian, does not
represent the actual distribution of the data. A natural alternative is SVM, a non-parametric
scheme based on the state-of-the-art statistical learning theory, to improve the classification
accuracy. SVM can also be applied to multispectral and hyperspectral images without suffering
from the “curse of high dimension”, or the so-called Hughes Effect. Remote sensing image
classification using SVM has been reported to be computationally simple and can result in better
accuracy compared to other more computationally intensive classifiers [19] – [21].