Buried Plant Detection: A Volterra Series Modelling Approach Using Artificial Neural Networks

This paper discusses the embedding of artificial neural networks (ANNs) into the framework of the Volterra series for modelling the problem of detecting buried pipes. This problem is formulated as a classification task whereby it is necessary to discriminate between the ground surface and an actual pipe reflection buried in noise in the return signal from ground probing radar. The objective is to filter out the unwanted surface reflection to enable improved mapping of the site being surveyed. Since the ANN correctly maps out a real test site, it can be viewed as having modelled the system transfer function relating the training patterns to their respective classes. Using the weights learnt by the ANN and its nodal functions, this transfer function is mathematically formulated. It is shown that the latter leads to a Volterra series representation of the pipe detection problem and effectively lends itself to the extraction of the Volterra kernels for this particular system. Copyright 1996 Elsevier Science Ltd