Improving Performance of Hyperspectral Unmixing Using Multi-Layer Perceptron by Generating an Appropriate Synthetic Library

Abstract

Hyperspectral unmixing (HSU) aims at extracting sub-pixel information and solving the mixed pixel problem in hyperspectral (HS) imaging. The multi-layer perceptron (MLP) network has been widely employed to address the problem thanks to its capability in learning the complex nonlinear relationship between the endmembers and their abundances. In this paper, we investigate the application of finding a suitable synthetic library for HSU using MLP. Firstly, fully constrained least square method is utilized to extract the initial solution because of its speed and high accuracy. The synthetic library is only made of samples around the initial solution, which avoids searching in the entire abundance range. Secondly, the discrete cosine transform and discrete wavelet transform are utilized to reduce the number of HS image bands in order to reduce the network's parameters resulting in preventing the overfitting. Since these feature reduction methods are data-independent, they are proper for compacting synthetic libraries. Furthermore, the spectral information divergence is utilized as an objective function in order to achieve a better result. The proposed method is applied to both synthetic and real datasets and the robustness and abundance estimation error has been evaluated. The results illustrate the potency of the proposed method in real applications.