Independent Component Analysis for coastal water mapping using hyperspectral datasets

Independent component analysis (ICA) is considered to be one of the most recent and successful ways to produce independent components out of the hyperspectral cube. The tool tries to resolve the blind source separation (BSS) statistical problem and has been applied to various case studies of hyperspectral datasets, for dimensionality reduction and separation of independent signal sources, i.e. endmembers. Many ICA algorithms have been proposed in the literature. In this study, the FastICA, JADE, BSS SVD, SONS, NG-OL, and SIMBEC algorithms were applied on airborne hyperspectral data for coastal water mapping. Emphasis was given on water turbidity. In order to enforce the capacities of FastICA, a methodology including the eigen-thresholding Harsanyi-Farrand-Chang noise suppression technique, as well as, three-level discrete wavelet transform (DWT) was developed. Results were compared and evaluated with in situ measurements related to turbidity. ICA algorithms produced quite interesting results. The BSS SVD algorithm was proven the most efficient tool for coastal water mapping.