Spectrally Segmented Linear Discriminant Analysis for Hyperspectral Image Classification

Abstract

Remote sensing Hyperspectral Image (HSI) contains notable information about the earth’s objects which is actually obtained by hundreds of closely spaced spectral bands. Due to the enormous size of the original HSI dataset and the large number of spectral bands, the expected classification performance is not achieved. As such, dimensionality reduction approaches using feature extraction and feature selection are generally adopted to enhance classification performance. The Principal Component Analysis (PCA) for unsupervised feature extraction, as well as the Linear Discriminant Analysis (LDA) for supervised feature extraction, commonly reduces linear features for feature subspace detection. However, as a consequence of taking the global structure of the feature, both PCA and LDA fail to extract the local structure of HSI. In this paper, we propose a spectrally segmented LDA (SSLDA) based feature extraction where we apply the LDA to each spectrally segmented subgroup of HSI data to extract better local characteristics in addition to global characteristics. Per-pixel classification using a Support Vector Machine (SVM) with RBF kernel is applied to our proposed SSLDA method, PCA, LDA, and spectrally segmented-PCA (SSPCA) on the Indian Pines agricultural HSI dataset. The experimental results illustrate that the overall classification accuracy of SSLDA (90.62%) remarkably performs better than all the other investigated methods: PCA (85.58%), LDA (86.46%), SSPCA (89.37%), and the complete original dataset without employing any feature reduction method (82.85%). The proposed SSLDA also requires the least amount of memory in different implementation scenarios.