An Evaluation of the Potential for Fuzzy Classification of Multispect ral Data Using Artificial Neural Networks

Fuzzy classification, or pixel unmixing, is the estimation of the proportion of the cover types from the composite spectrum of a mixed pixel. In this paper, we evaluate how the separation between class means, the covariance matrix of each class, and the relative location of the class means in the spectral space limit the fuzzy representation of mixtures. The influence of these factors is illustrated with a fuzzy classification using a back-propaga tion artificial neural net. Experiments using simulated data indicate that a fuzzy classification with an average error of less than 10 percent requires a Bhattacharyya Distance between classes of at least 9. The error in the fuzzy representation using a neural net also varies as the proportions of the classes changes, with a peak error when one class comprises approximately 0.20 to 0.25 of the mixed pixel. Back-propagation neural networks are not necessarily good at spectral unmixing. The backpropagation neural network produces spectral-space partitions between the classes that are generally steep, and that are not necessarily midway between the classes. The partitions tend to be simple, and somewhat linear. In addition, the output on nodes does not have to sum to 1.0, which may result in situations where high values are predicted for two classes simultaneously. Two methods of improving neural network behavior for fuzzy classification include use of a compound linear-sigmoid activation function, and training using synthetic mixed pixels.