Assessment of neural network input codings for classification of multispectral images

The research effort reported in this paper focuses on the evaluation of different input codings influencing the performance of a back-propagation neural network for classification of remotely sensed images. The clustering capability, which can be visualized through the Euclidean distance graph, is introduced as a tool to predict the credibility of the input coding. An investigation was also conducted to study the used of weight function to improve the clustering capability of the binary-coded-decimal input coding, a widely used coding approach in remote sensing area. Results obtained indicate that the classification performance of the neural network classifier is closely related to the clustering capability of the input codings. In order to fully exploit the generalization property of neural network, the clustering property of the classes must be maintained during the input coding process.