Segmented-Sparse-PCA for Hyperspectral Image Classification

Abstract

Remote sensing hyperspectral images (HSI) are typically acquired to obtain critical information about the land covers via adjacent cramped spectral wavelength bands. The classification performance appears inadequate when every original HSI band is in use. In order to improve performance of classification, band (dimensionality) reduction approaches using feature extraction as well as feature selection techniques are widely utilized to mitigate this. Despite the widespread usage of Principal Component Analysis (PCA) in HSI feature reduction, it frequently struggles to assess the local beneficial properties of the HSI as it analyzes solely the HSI's global data. Segmented-PCA (SPCA) and Sparse-PCA are initiatives to replace the PCA. In this paper, we propose the Segmented-Sparse-PCA (SSPCA) feature extraction approach nto exploit the amenities of both SPCA and Sparse- PCA. In particular, we first segment the entire dataset into a number of strongly correlated spectral band sub-groups and then apply Sparse-PCA to each subgroup separately. Afterward, the results are analyzed through experimenting over the mixed agricultural Indian Pines HSI dataset using a per-pixel Support Vector Machine (SVM) classifier. The experimental results exhibit that the overall classification performance of the proposed SSPCA (91.024%) performs better compared to every other investigated approaches: PCA (83.554%), SPCA (86.774%), Sparse-PCA (88.39%) and the entire original bands of HSI (82.997%).