Robust construction of radial basis function networks for classification

A neural network, based on robust construction of locally tuned radial basis functions (RBFs), is proposed to design a pattern classifier. A one-class one-network classification scheme is used to improve the data separation. A data sphering technique is applied to the raw training data for each class to decorrelate/normalize the data and to remove the potential outliers. The generalized Lloyd vector quantization clustering (LBG) algorithm with centroid splitting is applied on the sphered data to determine the centers and the diagonal covariance matrices of the Gaussian kernels. Better performance is achieved by the authors´ proposed method compared to an existing RBF construction method on artificial data. Favorable simulation results are achieved using the technique compared to other neural networks in classifying the Landsat-4 Thematic Mapper (TM) remote sensing data