Millions of acres of Department of Defense land (Formerly Used Defense Sites, sites designated for Base Realignment and Closure, and closed ranges on active installations) are potentially contaminated with unexploded ordnance (UXO) ranging in size from 20 mm shells to 2000 lb bombs. Current detection technologies exhibit limited capabilities in discriminating between hazardous ordnance items and nonhazardous clutter. Consequently, the false alarm rate induced by man-made and geological clutter is high and nearly 70 percent of the costs to remediate UXO sites are spent excavating non-ordnance items. The cost of identifying and disposing of UXO is estimated to range up to $100 billion using current technologies. Therefore, new technologies are needed to reduce the cost of disposal.
The objective of the project was to develop a single instrument combining target detection and accurate localization (via passive magnetic gradiometry of the earth’s-field distortion) with target-clutter discrimination (using the response to a broadband exciting field).
The instrument was to incorporate a full tensor magnetic gradiometer inside a three-axis excitation field source. With the source turned off, the gradiometer enables detection and localization. With the source turned on at closer range, the gradiometer measures the broadband response of the object and a classification algorithm would decide whether the object was a UXO or not.
Early on in the program, it was recognized that the passive measurement system component is more fully advanced than the multi-axis, broadband excitation component. In order to address the greater technical risk that the latter presents, the program was re-scoped to focus exclusively on the multi-axis, broadband system.
The researchers built and tested two wideband, multi-axis laboratory systems: a frequency domain (FD) operating over the frequency range 5 Hz – 100 kHz and a time domain (TD) system with the capability of measuring the full decay curve over the time window 10 μs – 2800 μs. Each has theoretical and practical advantages and drawbacks (particularly in the context of geophysical exploration systems), without a real resolution concerning their relative merits.
Data were collected on a larger sample of UXO and clutter items. All data were collected at ranges from the sensor to the test object that yielded a signal-to-noise ratio (SNR) of 5, as a fiducial standard. Data consisted of a training set—data collected from each of the types of UXO in known orientations, with ground truth—used to train the classification algorithms, and a test set—data collected from both clutter and UXO, in arbitrary orientations with the objects not identified to the classifiers.
Algorithms were trained on the training set and then tested against the test set. The models ranked the unknown test set from lowest to highest in order of likelihood of being UXO.
The FD results were not encouraging. A total of 27 objects, 8 UXO and 19 clutter, were tested with the FD system. If one simply divides the ranked list at the center into two segments and identifies the higher-likelihood segment as UXO and the lower-likelihood segment as clutter, one will have correctly identified exactly half the UXO and half the clutter items: the algorithm and data combine to give the result of a coin toss.
Initially, the result of the algorithm applied to the TD system data (using 24 objects, 7 UXO and 17 clutter) appeared even worse: only 1 UXO is identified correctly, and 6 are misidentified. However, simply inverting the output of the algorithm yields 6 of 7 UXO identified correctly, for a successful UXO identification rate of 85.7% and a successful clutter identification rate of 64.7%. Interpreted as conventional detection and false alarm probabilities, one obtains a detection rate of 85.7% at a false alarm rate of 35.3%.
Using funding separate from this project and not documented in this project’s final report, the researchers developed their own classification algorithms in the context of developing an advanced portal for concealed weapon detection. An initial look at that algorithm, suggests that the data provide more discriminatory power than the algorithms applied in this project were able to use. The concept of multi-axis, broadband response measurements may yet enable significant remediation cost reductions via improved target-clutter discrimination.