Target Detection in Hyperspectral Imagery Using Spectral Signature Matching Algorithms

This paper presents a method for detecting a known target (e.g., trees) in hyperspectral images using spectral matching techniques. The pure spectral signature of the target material is used to identify and locate the target within the image. Specifically, we employ the Spectral Angle Mapper (SAM) algorithm for spectral matching to detect the known target (trees). The reference spectrum for the trees is sourced from the ECOSTRESS spectral library and serves as the benchmark for spectral comparison. In this approach, the spectral signatures of each pixel in the hyperspectral data cube are compared with the reference spectrum. Pixels with the closest matching spectral characteristics are classified as part of the target material. An SAM score map is generated for different regions in the test dataset, quantifying the spectral match between each pixel's signature and the reference spectrum. Using a minimum score threshold, regions in the data are classified based on spectral similarity, assigning each pixel a class label according to its match with the reference spectra.The experimental results demonstrate that this thresholding approach successfully localizes target regions, resulting in a commendable mean squared error (MSE) of 0.020492 for tree labels, showcasing the algorithm's reliable detection performance. Furthermore, we compared the Spectral Angle Mapper (SAM) algorithm with other advanced spectral matching algorithms. It was found that the Normalized Spectral Similarity Score (NS3) algorithm achieved the lowest MSE of 0.018992, surpassing both SAM and the Jeffries-Matusita Spectral Angle Mapper (JMSAM).