Development of a new spectral library classifier for airborne hyperspectral images on heterogeneous environments

Abstract

The classification of hyperspectral images on heterogeneous environments without prior knowledge about the study area is a challenging task. Finding potential pure spectral signatures or endmembers (EM) of the various surface materials within an image is essential for obtaining accurate classification results. Automated endmember selection techniques, in many cases, return an unlabelled result without a relationship to a known material. This study demonstrates the potential of an automated spectral classification approach for hyperspectral imagery by using a comprehensive spectral library including a generalized class structure without the use of prior knowledge of the given scene. The classifier works by comparing every unknown image pixel to all labelled known spectra in the spectral library using a mixed measure similarity analysis of the spectral information divergence SID (Chang, 2000), the spectral angle mapper SAM (Kruse et. al., 1993) and the tangent trigonometric function (Du et. al., 2004). These similarity measures are the main criteria used to assign the class membership to a given pixel. In addition, a statistical analysis of the best ten scores identifies the statistical dominant material class from the similarity analysis. This statistical approach allows a pixel-related estimation of the classification reliability. The spectral library comparison classifier (SLC-Classifier) enables the classification of hyperspectral images on heterogeneous environments to be as complete as possible (depends on the input spectral library) with results containing both labelled potential pure spectra and spectra with low similarity agreement. Pixels with low similarity agreement are mixed pixels and pixels related to materials without good representative spectra in the comprehensive spectral library. We demonstrate that this classifier is suitable for the identification of surface materials using hyperspectral images were detailed knowledge about the environments does not exist.