Metric similarity regularizer to enhance pixel similarity performance for hyperspectral unmixing

Hyperspectral linear unmixing refers to the process that separates the pixels spectra from hyperspectral image into a collection of spectral signatures referred as endmembers and their abundances. In practice, the identified endmembers can vary spectrally within a given image and can thus construe as variable instances of pure endmembers. To this end, this work implements a linear unmixing system, which include noise estimation, band selection, and endmembers estimation. This work exploits the concept of lower ordered statistics to reduce within class variance while selecting the bands and maximum likelihood to observe the mixtures of spectral parameters, and expectation and maximization framework to estimate the mixture of parameters and to optimize the parameters. While optimizing the parameters, metric similarity regularizer incorporate to enforce the spatial related rationality and spectral similarity to make the system more effective. The main advantage of the system is, it can be used for both reduce and full wavelength data.