A method of endmember extraction in hyperspectral image based on landmark isometric mapping

A fast endmember extraction method based on landmark point selection is presented to overcome the high complexity and memory usage of the classical Isomap-NFINDR algorithm. The proposed method uses the maximin distance method to initial the k cluster centers, and carries out clustering segmentation using spectral angle instead of Euclidean distance. According to the spatial characteristics of the image, N landmark points which are near to cluster center are selected from the remaining points after removing the boundary points. Experiments with real images reveal that the algorithm proposed has the similar accuracy with the original algorithm and its operational efficiency is improved by 60 times.