Improved L-Isomap for classification of hyperspectral imagery via vector quantization

Landmark-Isometric mapping (L-Isomap) has greatly reduced the computational complexity of Isomap with the idea of landmarks. However, due to the irregular distribution of pixel points in spectral space, the usual random selected landmarks perform not well in hyperspectral image data. To solve the problem, Vector Quantization (VQ) has been introduced to improve the landmark selection. With two classifiers, the classification results of manifold coordinates from L-Isomap with VQ landmarks are compared with that of random landmarks. The results show that VQ landmarks could improve much the classification result in each class from random landmarks, and larger number of landmarks will lead to higher classification accuracy.