Pixel Unmixing in Hyperspectral Data by Means of Neural Networks

Neural networks (NNs) are recognized as very effective techniques when facing complex retrieval tasks in remote sensing. In this paper, the potential of NNs has been applied in solving the unmixing problem in hyperspectral data. In its complete form, the processing scheme uses an NN architecture consisting of two stages: the first stage reduces the dimension of the input vector, while the second stage performs the mapping from the reduced input vector to the abundance percentages. The dimensionality reduction is performed by the so-called autoassociative NNs, which yield a nonlinear principal component analysis of the data. The evaluation of the whole performance is carried out for different sets of experimental data. The first one is provided by the Airborne Hyperspectral Scanner. The second set consists of images from the Compact High-Resolution Imaging Spectrometer on board the Project for On-Board Autonomy satellite, and it includes multiangle and multitemporal acquisitions. The third set is represented by Airborne Visible/InfraRed Imaging Spectrometer measurements. A quantitative performance analysis has been carried out in terms of effectiveness in the dimensionality reduction phase and in terms of the accuracy in the final estimation. The results obtained, when compared with those produced by appropriate benchmark techniques, show the advantages of this approach.