Identification of most useful spectral ranges in improvement of target detection using hyperspectral data

Abstract

Hyperspectral imaging, because of its high spectral content, has been used in many surveillance and intelligence applications. Major issues in exploitation of HSI data, however, are spectral variability, noise, small size of the target and data dimensionality. Discrimination of target from background and/or other spectrally similar targets is also a key issue in target detection. Due to high spectral resolution, HSI data often contains huge amounts of redundant information which may adversely affect the performance of detection algorithms. Dimensionality reduction techniques are used to reduce this redundancy of HSI dataset but in case of target detection applications, the reduced/transformed dataset must preserve spectral bands/ranges containing important target information. Performing detection in spectral ranges containing most useful information about the target may improve its probability of detection. Several methods for identification of useful spectral ranges in detection have been reported in literature. The aim of this paper is to assess the applicability of Artificial Neural Network (ANN) in identifying the most useful spectral range for target discrimination. The Three fabric targets based on three different backgrounds have been analyzed using ANN to identify and the same has been further examined for its efficacy in improvement for target detection. ANN based analysis leads to two distinct spectral ranges which appear to improve target detection. © 2018 National Authority for Remote Sensing and Space Sciences