Evaluation of Dimensionality Reduction of Hyperspectral Images for Spatial-Spectral Classification Framework

Hyperspectral remote sensing facilitates detailed information about a ground scene or object of interest through three-dimensional hyperspectral images. These images are essentially composed of hundreds of two-dimensional image bands, which are the accumulation of spectral reflectance from visible to infrared wavelength range. These hyperspectral images potentially hold a wealth of spectral and spatial information that enables fine separation among similar surface objects. However, these high-dimensional datasets also contain redundant and correlated information, which brings several challenges during the classification process. Therefore, dimensionality reduction becomes a crucial preprocessing step for extracting relevant and compact information. The present study aims to evaluate various feature extraction based dimensionality reduction (DR) techniques such as Independent Component Analysis (ICA), Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), and kernel Principal Component Analysis (kPCA) for spectral-spatial classification framework. These techniques are employed to extract meaningful and compact spectral features. However, spatial filters such as the Gabor filter, Entropy Filter, Standard Deviation Filter, and Range Filter are utilized to extract spatial features. The spectral and spatial features are merged together for Support Vector Machine (SVM) based classification. The experimentation has been performed with Indian Pines and Pavia University Scene hyperspectral datasets. The results demonstrate that the PCA technique outperformed others in terms of overall accuracy for these datasets.