Hyperspectral data classification using an ensemble of class-dependent neural networks

Hyperspectral data are characterized by a huge size due to hundreds of narrow frequency bands. However, the classes of interest are often characterized by only a few features from the available (ormodified) feature space. Using a few samples of the classes of interest it is possible to identify the features characterizing the classes by calculating the Bhattacharya distance, B or the Jeffries-Matusita distance, J for every feature and for every class combination. However, the classification using these features is not trivial. We use a new architecture based on class-dependent neural networks for this purpose. Class-dependent neural network is a feed forward neural network for every class with features characterizing only that class as inputs. In the combined architecture, all the classes are first, individually separated from other classes using first level class-dependent neural networks to map the characteristic features to a fuzzy value for each of the classes. Then, a final classification decision is made using a second level neural network with inputs from the outputs of the first level neural networks.