Computationally-Efficient Bandwise GBM Model for Hyperspectral Nonlinear Unmixing

Nonlinear Unmixing using the Band-wise Generalized Bilinear Mixing (NU-BGBM) model specifies an acceptable mixing scenario up to the second-order interaction of light rays and also suppresses various types of mixed noise while performing un-mixing. However, NU-BGBM requires high computational time and multiple parameter tuning, which could practically limit its application to large HyperSpectral Images (HSIs). In this context, we propose a computationally efficient BGBM as a fast and robust variant of the NU-BGBM method, In this model, the objective function for the non-linear optimization scheme is designed without the sparsity constraint, and an iterative scheme based on the Alternating Direction Method of Multipliers (ADMM) is formulated for solving the proposed model. Extensive analyses have been carried out on synthetic (with simulated mixed noise) and real HSIs. The performance of the proposed method was compared with the NU-BGBM model using signal-to-reconstruction error (SRE), abundance Root-Mean-Square Error (aRMSE), source Root-Mean-Square Error (sRMSE), and Root-Sum-Squared (RSS) error. Results from extensive numerical analysis reveal that the proposed method reduces computation time (on an average six times faster) while being comparable (and often better) than NU-BGBM in terms of accuracy on large data sets.