Impacts of Noise on the Accuracy of Hyperspectral Image Classification by SVM

The support vector machine (SVM) has become a popular tool for image classification recently. The performance of SVM for hyperspectral image classification has been examined from a range of perspectives, but the impacts of noise, errors and uncertainties have attracted less attention. This paper aims to evaluate the impacts of noise on SVM classification. The research is undertaken using real imagery acquired by the OMIS hyperspectral sensor. To assess the sensitivity and reduction capacity of SVM classifier to different types of noise a simulation study is undertaken using two types of noise. The first type of noise is striping, in which some rows or columns of the image have markedly abnormal signals. The second type of noise is caused by some uncertain factors that may impact upon one band, one pixel or one line. This noise may be evaluated by introducing salt and pepper noise. A variety of datasets containing different types of noise are generated and classified using a SVM. The results of the classifications, with particular regard to their accuracy, are compared against a classification of the original dataset and comparative analyses obtained using traditional classifiers including the spectral angle mapper (SAM) and binary encoding (BE). The results indicate that the SVM is more effective to alleviate the effects of noise than SAM and BE.