FastICA(MNF) for feature generation in hyperspectral imagery

The improvement in sensor technologies over the recent years is providing the earth observation community with datacubes of several hundreds of spectral bands which are both an incredible opportunity for phenomenology understanding and material characterization but also pose a serious challenge for their exploitation. We propose in this paper to eliminate spectral redundancy and noise with the minimum noise fraction (MNF) transform followed by the extraction of statistical independent components using FastlCA. This processing is applied successively on the 0.4 to 2.4 mum spectrum and on spectral domains of similar 2nd order statistics (VIS, VIS+SWIR, SWIR) from a scene collected by the Hyperion sensor. Results show that specific features are generated in each of these domains that are not necessarily captured when executing the processing on the whole spectrum.