Comparison of Feature Selection Techniques for SVM Classification

The use of satellite imagery in the derivation of land cover information has yielded immense dividends to numerous application fields such as environmental monitoring and modeling, map making and revision and urban studies. The extraction of this information from images is made possible by various classification algorithms each with different advantages and disadvantages. Support Vector machines (SVMs) are a new classifier with roots in statistical learning theory and their success in fields like machine vision have drawn the attention of the remote sensing community. Previous studies have focused on how SVMs compare with traditional classifiers such as maximum likelihood and minimum distance to means classifiers. They have also been compared to newer generation classifiers such as decision trees and artificial neural networks. In this research the understanding of the application of SVMs to image classification is furthered by proposing feature selection as a way in which remote sensing data can be optimized. Feature selection involves selecting a subset of features (e.g. bands) from the original set of bands that captures the relevant properties of the data to enable adequate classification. Two feature selection techniques are explored namely exhaustive search and population based incremental learning. Of critical importance to any feature selection technique is the choice of criterion function. In this research a new criterion function called Thornton’s separability index has been successfully deployed for the optimization of remote sensing data for SVM classification.