An Improved Endmember Extraction Algorithm by Inversing Linear Mixing Model

In hyperspectral imagery there are some cases when no pure pixels present due to the limitation of the sensors’ space resolution and the complexity of the ground components, and then the endmembers extracted by traditional algorithms are usually mixing ones still. In order to solve this problem, this paper proposes an endmember extraction algorithm based on the re-analysis of preliminary endmembers extracted by volume calculating concept under the linear mixing model. After extracting the pixels which are most approximated to the pure pixels from the image, using the convex polyhedron’s geometric characters to search out the boundary pixels which are around the preliminary endmembers and on the edge of the convex polyhedron formed by the pure pixels. Calculating the abundance fractions of every endmember in these pixels by the laws of sins, thus, with these coefficients the endmembers could be obtained using the inversion of linear mixing model. Hyperspectral scenes are simulated by the real spectra to investigate the performance of the algorithm. Preliminary results indicate the effectiveness of the algorithm. Applying the algorithm to a real Hyperion scene it also gets a better result.