A neural network approach for pixel unmixing in hyperspectral data

Neural networks algorithms have already shown good capabilities in handling nonlinear inversion problems in hyperspectral remote sensing. In this study we investigate on their potential in solving spectral unmixing. A Multi-Layer Perceptron (MLP) neural network scheme is trained for the implementation of a pixel-based classification algorithm. Subsequently, for the output response, the “winner-takes-all” rule is replaced by a more soft interpretation able to give the percentage with which, each of the considered land cover classes, may be associated to the analysed pixel. In an experimental set-up addressing multi-temporal and multi-angular CHRIS-PROBA imagery, the results obtained with such a technique have been compared with those yielded by Linear Spectral Unmixing (LSU), up to date one of the most frequently used approach for dealing with the unmixing problems.