Feasibility Study on Using Physics-Based Modeler Outputs to Train Probabilistic Neural Networks for UXO Classification

摘要

A probabilistic neural network (PNN) has been applied to the detection and classification of unexploded ordnance (UXO) measured using magnetometry data collected using the Multi-sensor Towed Array Detection System (MTADS). Physical parameters obtained from a physics based modeler were used to describe the UXO and scrap targets found at three sites: Badlands Bombing Range (BBR) Target 1 and 2 and the Former Buckley Field. The PNN was trained and tested using cross validation (CV) software developed at NRL. The PNN was able to correctly identify between 84% to 94% of the targets. By adjusting the probability threshold, further improvements in the discrimination of UXO were possible: 96% of the UXO were correctly identified for BBR Target 1, 100% for BBR Target 2, and 94% for the former Buckley Field. The ability to train using one site (BBR target 2) and predict another (BBR Target 1) was successful with 95% of the UXO correctly identified and a false alarm rate of 35%.