An investigation of methods for class mean and covariance initialization of a stochastic mixing model for hyperspectral imagery is described along with other relevant issues concerning algorithm convergence such as updating of the class priors, constraining the mixture classes and the number of fraction levels and endmember classes. The various refinements of the iterative algorithm are presented and tested on synthetically-generated test data as well as real reflective hyperspectral imagery, and recommendations are made concerning how the stochastic mixing model can be best implemented. The results show that the refined stochastic mixng model is a robust approach for unmixing hyperspectral imagery with different levels of complexity.

Date Published: 7 January 2004
PDF: 12 pages
Proc. SPIE 5159, Imaging Spectrometry IX, (7 January 2004); doi: 10.1117/12.499680

Published in SPIE Proceedings Vol. 5159:
Imaging Spectrometry IX
Sylvia S. Shen; Paul E. Lewis, Editor(s)