Design and Development of Higher Order Spectral Unmixing Model for Mangrove Species Discrimination

Abstract

The aim of this research is to extract the potential of hyperspectral imagery acquired by the hyperion sensor for improved species level discrimination of pure and mixed mangrove patches growing in natural habitats. The radiations that emanate from these pure and mixed mangrove stands are both linear and nonlinear in nature that may be suitably modeled for accurate endmember detection and abundance estimation. While linear interactions are usually taken care of by extant linear spectral unmixing models, it is difficult to compute the intricate cobweb of non-linear interactions resulting out of pure and mixed pixels. The closed-patch mangrove forest of Sunderban are particularly characterized by mixed stands (pixels) that bears several mangrove species (endmembers) embedded in close system. In order to fully characterize these non-linear interactions, a new ‘higher order non-linear spectral unmixing model’ has been developed in this study. This new model helps to overcome the limitations of the prevalent linear spectral unmixing models that take care of single interactions. The developed model further leads to better characterization of intra-species interactions between similar mangrove species of pure stands and gives rise to more accurate fractional abundance data than the linear models. Attempt has been made to apply the developed non-linear model at several key mangrove patches of Sunderban. The application proved highly successful in identification of admixtures of mangrove endmembers comprising Excoecariaagallocha, Ceriopsdecandra, Avicennia marina and Phoenixpaludosa.