Successful discrimination between unexploded ordnance (UXO) and clutter depends on the existence of characteristic attributes associated with UXO that are consistent across all examples of an ordnance type and that distinguish them from clutter. However, real UXO are often dented, bent, broken, or otherwise damaged. This produces variability in their electromagnetic induction (EMI) response and complicates the problem of UXO/clutter discrimination. The degree of consistency from one item to the next within a given class or sub-class of UXO is related to the peak theoretical performance of any discrimination scheme used to classify targets, and this represents a fundamental limit to the performance of all discrimination schemes. This project was aimed at quantifying UXO variability by measuring a large number of real UXO and expressing variability in physically meaningful terms. Similar to navigation errors, modeling errors, and sensor noise, inherent UXO variability degrades discrimination performance, but unlike these other sources of error, it cannot be reduced through better engineering or better processing tools.
The overall objective of this project was to promote improvement of discrimination algorithms by quantifying the inherent variability of UXO themselves. This information is useful for transitioning research results, which are often obtained under controlled test environments utilizing small numbers of representative samples, to complex and diverse real-world settings.
Researchers collected EMI data on 664 real UXO items from four Department of Defense sites using two state-of-the-art sensors: GEM-3 (frequency-domain) and EM-63 (time-domain). For each UXO type, the researchers aimed to collect samples of sufficient size so that variability in the population could be quantified with reasonable accuracy. From 10 to 30 items was usually sufficient, as determined through Jackknife and Bootstrap resampling. Analysis was done by fitting each target individually using a joint time-domain - frequency-domain (TDFD) model, and four physical parameters were derived from the fits: (1) magnitude, (2) fundamental decay constant, (3) magnetic crossover time, and (4) static dipole strength. Mean and variance of these parameters were found for each UXO class, and all data was grouped and summarized for each class.
Results provided by this program, combined with ancillary ongoing and future research, provide the user community with the tools and information required to reduce false alarms while maximizing detections. (Project Completed - 2007)