Unmixing of hyperpectral image using nonlinear parametric model

Abstract

Hyperspectral image contains hundreds of bands so it is spectrally overloaded and contains extent information to differentiate spectrally unique material. Hyperspectral data generally used to identify the presence of material in scene. Almost all the hyperspectral cameras have spatial resolution limit (>5m per pixel) due to that each pixel can be a mixture of several materials. The process of unmixing is to unmixone of these mixed pixels. There are two models available to approximate mixing, (i) Linear Mixing Model (LMM) (ii) Nonlinear Mixing Model (NMM). Over a time, various approaches have been devised to address LMM and it's unmixing. In LMM, macrospectral mixtures are assumed. Nonlinear model comes under consideration due to microscopic mixing scale. In this paper, Generalized bilinear model is used which is nonlinear parametric model to get mixed data. Its accuracy depends on parametric form and parameter value chosen. It comes under convex optimization problem, so it can be solved using any optimization technique. Gradient descent algorithm (GDA) is employed to solve this optimization problem. Advantage of GDA over other unmixing techniques is that it transforms nonlinear model into linear one. To improve unmixing result, it is indeed advisable to consider spatial correlation among abundances. A novel approach has been introduced in this paper which considers 2nd order neighborhood correlation between abundances. Using our approach one can achieve better segmentation.