Convex Geometry Based Endmember Extraction for Hyperspectral Images Classification
Nian Zhang, Wagdy Mahmoud
Abstract
Hyperspectral imaging, initially developed for defense purposes, has found widespread applications across diverse fields such as food safety, agriculture, environmental monitoring, intelligence, and medical imaging. These images consist of numerous spectral bands with narrow bandwidths, enabling precise analysis. The primary task involves identifying “endmembers,” which are pure signatures defining distinct spectral classes within the data. This paper investigates the most popular convex geometry-based endmember extraction algorithms and then classifies or identifies materials using spectral matching in hyperspectral images. Specifically, we first perform endmember extraction, in which pure spectral signatures are identified in the hyperspectral image. These signatures serve as reference points for different classes of materials. Then spectral matching method is adopted to not only find the spectral similarity between the ground truth endmembers and the extracted endmembers, but also compare the spectral signatures of pixels in the image to the extracted endmember spectra. A score map is generated for various regions in the test data. This map is created by calculating the spectral match score between each pixel's spectrum and the pure reference spectrum. Regions in the test data are classified based on a minimum score criterion. Each pixel in the test data is assigned a class label according to the similarity of its spectrum with the reference spectra. The pure spectra used in this paper are sourced from the ECOSTRESS spectral library. The experimental outcomes reveal that the N-FINDR, PPI, and FIPPI methods can all identify the endmember signatures. Among them, the pixel purity index (PPI) method demonstrates the highest classification accuracy, while the fast iterative pixel purity index (FIPPI) approach takes the shortest execution time on hyperspectral images.