Estimating the Number of Endmembers in Hyperspectral Imagery Using Hierarchical Agglomerate Clustering

A classical spectral un-mixing of hyperspectral image involves identifying the unique signatures of the endmembers (i.e. pure materials) and estimating the proportions of endmembers for each pixel by inversion. The key to successful spectral un-mixing is indicating the number of endmembers and their corresponding spectral signatures. Currently, eigenvalue-based estimation of the number of endmembers in hyperspectral image is widely used. However, the eigenvalue-based methods are difficult to separate signal sources such as anomalies. In this paper, a two-stage process is proposed to estimate the endmember numbers. At the preprocessing stage, the spectral dimensions are reduced using principal component analysis and the spatial dimensions are reduced using convex hull computation based on reduced-spectral bands. At the hierarchical agglomerate clustering stage, a pixel vector is found by applying orthogonal subspace projection (OSP) and cluster pixel vectors using the spectral angle mapper (SAM), hierarchically. If the number of pixel vectors in a cluster is greater than the predefined number, the found pixel vector is set as the endmember. Otherwise, anomalous vectors are found. The proposed method was carried with both synthetic and real images for estimating the number of endmembers. The results demonstrate that the proposed method can be used to estimate more reasonable and precise number of endmembers than the eigenvalue-based methods.