Application of non-linear spectral unmixing on hyperspectral data for species level classification of mangroves

The present study is the first attempt to make use of hyperspectral data in the Sunderban eco-geographic province to enable species level discrimination of mangroves. Our objective here, is to unmix hypespectral images using non-linear spectral unmixing techniques by taking into account the higher order interactions of light that occurs among different target endmembers (mangrove species). The linear mixing models have also provided a good abstraction of the mixing process, but some naturally occurring situations exist where nonlinear models provide the most accurate assessment of endmember abundance. It has been noted that the nonlinear models have been successful in estimating the abundances for the different endmembers in places where the non-linear situation is prevalent within the mangrove forests with several layers of tree canopy considered to be present one above the other. For such a situation, the second order interaction among the endmembers have been considered. This paper applies and compares the classification accuracy of non-linear techniques using the two methods shown in the Bilinear Model. They are Nascimento´s and Fan´s Bilinear model for unmixing hyperspectral images. On comparison, Fan´s model has been able to classify mixed mangrove species more accurately than Nascimento´s model. It has been possible to identify 7 dominant pure and mixed mangrove species present in the study area.