Dimensionality reduction methods for hyperspectral image classification

Abstract

The application of hyperspectral imaging in various fields is growing rapidly, especially in the field of image classification since different spectral bands are available for the acquisition of an image. In this work, two different image types (rural and urban areas) are used for the implementation of hyperspectral image classification, where the main objective is to reduce the number of bands and training samples, and improve the accuracy. Both spectral and spatial features are considered.To perform dimensionality reduction of the original hyperspectral data three selection methods are applied. The first one is the Principal Component Analysis(PCA), the two remaining methods are Kernel PCA and Reduced KPCA which are both resulting from the improvement of principal component analysis. To extract spatial features, morphological filters are selected depending on the appropriate parameters that suit each image. The spectral and spatial features are then combined and the obtained data are presented to the classification algorithm that is the support vector machine (SVM) in this project.For Airborne Visible Infrared Imaging Spectrometer (AVIRIS), which is a rural area, morphological PCA is the one that gives the highest accuracy (96.8192). However, morphological Kernel PCA (KPCA) and morphological Reduced KPCA achieve a better result on Pavia Center and Pavia University which are urban areas, the accuracies are 97.4655 and 98.5338, respectively, for Pavia Center. The results obtained by different dimensionality reduction methods turned out to be promising.