Camp Butner Live-Site UXO Classification Using Hierarchical Clustering and Gaussian Mixture Modeling

Abstract

We demonstrate in detail a semisupervised scheme to classify unexploded ordnance (UXO) by using as an example the data collected with a time-domain electromagnetic towed array detection system during a live-site blind test conducted at the former Camp Butner in North Carolina, USA. The model that we use to characterize targets and generate discrimination features relies on a solution of the inverse UXO problem using the orthonormalized volume magnetic source model. Unlike other classification techniques, which often rely on library matching or expert knowledge, our combined clustering/Gaussian-mixture-model approach first uses the inherent properties of the data in feature space to build a custom training list that is then used to score all unknown targets by assigning them a likelihood of being UXO. The ground truth for the most likely candidates is then requested and used to correct the model parameters and reassign the scores. The process is repeated several times until the desired statistical margin is reached, at which point a final dig is produced. Our method could decrease intervention by human experts and, as the results of the blind test show, identify all targets of interest correctly while minimizing false-alarm counts.