Hyperspectral unmixing based on ISOMAP and spatial information

Several nonlinear techniques have recently been proposed for classification and unmixing applications in hyperspectral image processing. A commonly used data-driven approach for treating nonlinear problems employs the geodesic distances on the data manifold as the property of interest. Although this approach often produces better results than linear unmixing algorithms, the graph-based method treats an image as a bag of spectral signatures and ignores the relationship between the pixel and its spatial neighbors. To utilize the spatial distribution of pixels and improve hyperspectral unmixing precision effectively, a new method is proposed for incorporating nonlinear dimension reduction and spatial information, using isometric mapping (ISOMAP) to find significant low-dimensional structures hidden in high-dimensional hyperspectral data. Spatial information is also introduced into the traditional spectral-based endmember search process. A fully constrained least-squares algorithm is used to evaluate the abundance of each endmember. The experimental results for actual images reveal that the performance of the proposed method obtains much better unmixing results than the classical N-FINDR and ISOMAP algorithms.