Discrimination of Unexploded Ordnance from Clutter Using Linear Genetic Programming

Abstract

We used Linear Genetic Programming (LGP) to study theextent to which automated learning techniques may beused to improve Unexploded Ordinance (UXO)discrimination from Protem-47 and Geonics EM61non-invasive electromagnetic sensors. We conclude that:(1) Even after geophysicists have analysed the EM61signals and ranked anomalies in order of the likelihoodthat each comprises UXO, our LGP tool was able tosubstantially improve the discrimination of UXO fromscrap preexisting techniques require digging 62percentmore holes to locate all UXO on a range than do LGPderived models; (2) LGP can improve discrimination eventhough trained on a very small number of examples ofUXO; and (3) LGP can improve UXO discrimination on datasets that contain a high-level of noise and littlepreprocessing.