Classification of very high dimensional data using neural networks

Neural networks that are used in the classification of very-high-dimensional remotely sensed data are discussed in comparison to statistical classification methods. Both two-layer and three-layer iterative neural networks are used in experiments together with a parallel hierarchical neural network. The statistical methods applied include the maximum likelihood method, the minimum Euclidian distance, and two pooling methods (statistical multisource classification and the linear pool). The data used in experiments are simulated High-Resolution Imaging Spectrometer (HIRIS) data. All the methods are compared based on classification performance with different numbers of features, different numbers of training samples, speed (CPU time), and classification accuracy for training and test data. Statistical methods are shown to yield performance superior to that of the neural network methods.