This paper discusses two novel artificial neural network architectures applied to multi-class classification problems of remote-sensing data. These approaches are 1) a spiking-neural-network model for the partitioning of data into clusters, and 2) a neuron model based on complex-valued weights (CVN). In the former model, the learning process is based on the Spike Timing-Dependent Plasticity rule under the Hebbian Learning framework. With temporally encoded inputs, the synaptic efficiencies of the delays between the pre- and post-synaptic spikes can store the information of different data clusters. With the encoding method using Gaussian receptive fields, the model was applied to the remote-sensing data. The result showed that it could provide more useful information than using traditional clustering method such as K-means. The CVN model has proved to be more powerful than traditional neuron models in solving the XOR problem and image processing problems. This paper discusses an implementation of the complex-valued neuron in NRBF neural networks to improve the NRBF structure. The complex-valued weights are used in the supervised learning part of an NRBF neural network. This classifier was tested with satellite multi-spectral image data and results show that this neural network model is more accurate and powerful than the conventional NRBF model.
This paper describes a novel classification technique-NRBF (Normalized Radial Basis Function) neural network classifier based on spectral clustering methods. The spectral method is used in the unsupervised learning part of the NRBF neural networks. Compared with other general clustering methods used in NRBF neural networks, such as KMeans, the spectral method can avoid the local minima problem and therefore multiple restarts are not necessary to obtain a good solution. This classifier was tested with satellite multi-spectral image data of New England acquired by Landsat 7 ETM+ sensors. Classification results show that this new neural network model is more accurate and robust than the conventional RBF model. Furthermore, we analyze how the number of the hidden units affects training and testing accuracy. These results suggest that this new model may be an effective method for classification of multispectral
satellite image data.