Classification of multispectral images using BP-neural network classifier-input codings assessment

The research effort reported in this paper focused on the evaluation of different input codings influencing the performance of a backpropagation (BP) neural network for the classification of multispectral images. The assessments of the input codings are based on the performances of a network classifier using five different input coding schemes, namely normalization, temperature, coarse, binary coded decimal and Gray codings. The clustering property, which can be visualized through the “Euclidean distance” graph, is also introduced as a tool to predict the generalization capability of each input coding method. Experimental results obtained indicated that in order to fully exploit the generalization property of the neural network, the clustering property of the spectral features must be maintained during the input coding process