Hyperspectral Nonlinear Unmixing: Endmember Extracting Using Iterative Simplex CNN Method

摘要

Hyperspectral remote sensing images are showing rapid improvements in many domain such as weather, land cover classification, underwater species identification, exploring the space etc. Over the past decades pixel-wise classification have been focussed by the research community by improving accuracy, but still the problems persists due to low spatial resolution and multiple scattering effects of light generates the mixed substances in a pixel. In recent decade, the sub-pixel-wise and object based methods are focussed in a hyperspectral image classification. This paper concentrates on the experimentation of hyperspectral sub-pixel wise classification. There are two different types of spectral unmixing techniques are based on linear and nonlinear unmixing approaches. In linear unmixing techniques assumes that in macroscopic pixel level, pure substances may present per pixel, but in reality the intimate mixtures present in microscopic level due to the scattering effects of light. Basically there are two step involved in the nonlinear unmixing: endmembers extraction and abundances fractions present per pixel. This paper proposed the novel iterative simplex volume analysis Convolutional Neural Network (IS_CNN) method to extract the end members. And further, this research work employed fully constrained least square (FCLS) method to extract the endmembers of the hyperspectral data. The overall performance of the proposed method is measured with Root mean squared error (RMSE), the IS_CNN RMSE Average value for all extracted end members is 0.08 which shows the significant performance when compared with the FCLS method.