Dependent component analysis as a tool for blind spectral unmixing of remote sensed images

In this work, we present a blind technique for the estimation of the material abundances per pixel (end-members) in hyperspectral remote-sensed images. Classical spectral unmixing techniques require the knowledge of the existing materials and their spectra. This is a problem when no prior information is available. Some techniques based on independent component analysis proved not to be very efficient for the strong dependence among the material abundances always found in real data. We approach the problem of blind separation of end members by applying the MaxNG algorithm, which is capable to separate even sensibly dependent signals. We also present a minimum-mean-squared-error method to estimate the unknown scale factors by exploiting the source constraint. The results shown here have been obtained from either synthetic or real data. The synthetic images have been generated by a noisy linear mixture model with real, spatially variable, endmember spectra. The real images have been captured by the MIVIS airborne imaging spectrometer. Our results showed that MaxNG is able to separate the endmembers successfully if a linear mixing model holds true and for low noise and reduced spectral variability conditions.