An Expert-Free Technique For Live Site UXO Target Classification

In this paper we examine methods of automatic classification applied to Unexploded Ordnance (UXO) across data sets from a live site. All sensors used are time-domain Electromagnetic Induction (EMI) sensors. We solve for target extrinsic and intrinsic parameters using the Differential Evolution (DE) and Ortho-Normalized Volume Magnetic Source (ONVMS) algorithm. This inversion provides target locations and intrinsic time-series total ONVMS principal eigenvalues. We fit these to an empirical power decay model, the Pasion-Oldenburg model, providing dimensionality reduction for a Machine Learning (ML) approach. We group anomalies by the unsupervised Weighted-Pair Group Method with Averaging (WPGMA) algorithm. After requesting Ground Truths (GT) for the central element of each cluster, we train a supervised Gaussian Mixture Model (GMM), in which each class of UXO is represented by a multivariate Gaussian probability density. We request Ground Truths in rounds until we are confident there are no remaining Targets of Interest (TOI) in our survey of the site. Our system for UXO cleanup is fully automatic and expert free, and uses a priori knowledge combined with a learned algorithm.