Spectrally-Segmented-Incremental-PCA for Hyperspectral Image Classification

Abstract

Remote sensing through neighboring constrained spectral wavelength bands, the hyperspectral image (HSI) contains significant information about the land objects. Using all of the original HSI features (bands), it appears that the classification performance is inadequate. To attenuate this, band (dimensionality) reduction schemes using feature extraction and feature selection techniques are frequently used in order to enhance classification performance. Despite being often employed for HSI feature reduction, Principal Component Analysis (PCA) usually struggles to retrieve the local desired HSI features since it only evaluates the HSI’s global statistics. Therefore, Spectrally-Segmented-PCA (SSPCA) and Incremental-PCA (IPCA) are presented to supplant the classical PCA. In this paper, we propose the Spectrally-Segmented-Incremental-PCA (SSIPCA) feature extraction approach to make use of the utility of both the SSPCA and the IPCA. Specifically, SSIPCA divides the whole HSI into a number of spectrally separated bands’ subgroups before applying the standard IPCA to each subgroup independently. We experiment with the Indian Pines mixed agricultural HSI classification to assess the proposed SSIPCA employing a perpixel Support Vector Machine (SVM) as the classifier. Based on the classification accuracy, we evince that the proposed SSIPCA approach (90.78% & 88.702%) outperforms the entire original bands of HSI (87.610% & 86.361%), PCA (88.78% & 86.985%), IPCA (89.171% & 86.576%) and SSPCA (90.634% & 88.468%) feature extraction methods.