Sensitivity Analysis of Support Vector Machine in Classification of Hyperspectral Imagery

Nowadays by developing hyperspectral sensor technology, it is possible to simultaneously capture image with hundreds of contiguous narrow spectral bands. Increasing spectral bands provide more information and seem to improve classification accuracy. Nevertheless limited training samples lead to poor parameter estimation of statistical classifiers which is called Hughes phenomena. Recently Support Vector Machines (SVMs) are applied successfully for classification of hyperspectral imagery because they characterize classes by a geometrical criterion, not by statistical criteria. However, accuracy and performance sensitivity of SVMs in classification of hyperspectral imagery are affected by three different factors. The first one is the type of input data space which can be spectral space or feature space. In this paper three feature extraction methods, include: Principle Component Analysis (PCA), Independent Component Analysis (ICA) and Linear Discriminate Analysis (LDA) are used. Another effective factor is spectral similarity measures. Most of studies use Euclidean distance as a metric for measuring similarity between samples. By using Euclidean distance, geometric behaviour of data is evaluated and spectral meaning is not considered. This paper evaluates the effect of different metrics such as Spectral Angle Mapper (SAM) and Spectral Information Divergence (SID) on accuracy of classification. The last factor is training sample size that effect of this factor on SVMs classification accuracy is evaluated and results were compared with K-Nearest Neighbour (KNN) classifier. For evaluating sensitivity analysis of SVMs respect to these factors, polynomial and Gaussian kernels and two usual multiclass classification strategies include one against one and one against all are applied. Also experiments are carried out on the AVIRIS dataset. * Corresponding author