Significant progress has been made in unexploded ordnance (UXO) discrimination technology development. To date, testing of these approaches has been primarily limited to test sites with only limited application at live sites. Acceptance of discrimination technologies requires demonstration of system capabilities at real UXO sites under real world conditions. Any attempt to declare detected anomalies to be harmless and requiring no further investigation will require demonstration to regulators of not only individual technologies but an entire decision making process.
The principal objectives of the work conducted in this project were to develop a practical strategy for discrimination of UXO that can be reliably applied to real sites along with the protocols and tools to test performance. Three different demonstrations were conducted under this project.
The first demonstration of the methodology was conducted at the Former Lowry Bombing and Gunnery Range (FLBGR) in Colorado during the 2006 field season. The focus of the FLBGR demonstration was on verification of the single inversion process used to extract physics-based parameters from magnetic and electromagnetic induction (EMI) anomalies, and on the statistical classification algorithms used to make discrimination decisions from those parameters. Two sites were visited at FLBGR. The Rocket Range (RR) survey objective was to discriminate a mixed range of projectiles with a minimum diameter of 37 mm from shrapnel, junk, 20 mm projectiles, and small arms. The 20 mm Range-Fan (RF) survey presented a small item discrimination scenario with a survey objective of discriminating 37 mm projectiles from ubiquitous 20 mm projectiles and 50-caliber bullets.
The second demonstration was conducted as part of the ESTCP discrimination pilot study at Camp Sibert, Alabama, during 2007. The objective was to find potentially hazardous 4.2-inch mortars. The demonstration provided another test of the methodology as well as that of the cooperative inversion process. Both cued interrogation and full coverage data collected by different demonstrators were analyzed, allowing the effect of data quality on discrimination decisions to be assessed. For the Camp Sibert discrimination study, the project team created eight different dig-sheets from six different sensor combinations: (1) Multi-sensor Towed Array Detection System (MTADS) magnetics; (2) EM61 cart (classification and size based); (3) MTADS EM61 (classification and size based); (4) MTADS EM61 and magnetics; (5) EM63; and (6) EM63 and magnetics.
The objectives of the third demonstration were to evaluate the discrimination potential of the Geonics EM63 at Fort McClellan, Alabama, when deployed in a cued interrogation mode. Pasion-Oldenburg polarization tensor models were fit to each of the EM63 cued anomalies. Ground truth information from 60 of the 401 live-site anomalies, along with 18 items in the geophysical prove-out and 21 items measured in a test pit were used to train a statistical classifier. Features related to shape, encapsulated in the relative values of the primary, secondary, and tertiary polarizations, were unstable and could not be used for reliable discrimination. A feature space comprising the size and the relative decay rate of the primary polarization was used for discrimination of the medium caliber projectiles (75 mm and 3.8-inch shrapnel).
At the FLBGR sites, and for both instruments, the support vector machine (SVM) classifier outperformed a ranking based on amplitude alone. In each case, the last detected UXO was ranked quite high by the SVM classifier, and digging to that point would have resulted in a 60-90% reduction in the number of false alarms. This operating point is of course unknown prior to digging. The demonstrators found that using a stop-digging criteria of f=0 (midway between UXO and clutter class support planes) was too aggressive and more excavations were typically required for full recovery of detected UXO. Both the amplitude and SVM methods performed quite poorly on two deep (40 cm) emplaced 37 mm projectiles at the 20 mm RF, exposing a potential weakness of the goodness-of-fit metric. Retrospective analysis revealed that thresholding on the size of the polarization tensor alone would have yielded good discrimination performance.
At Camp Sibert, the results for all sensor combinations were excellent, with just one false negative for the EM63 when inverted without cooperative constraints. When inverted cooperatively, the EM63 cued interrogation was the most effective discriminator. All 33 UXO were recovered with 25 false alarms (16 of these were in the “can’t-analyze” category). Not counting the can’t-analyze category, the first 33 recommended excavations were all UXO. The MTADS and MTADS cooperatively inverted were also very effective at discrimination, with all UXO recovered very early in the dig list (e.g., for the MTADS cooperative there were just two false positives (FP) by the time all 117 can’t-analyze UXO were recovered). The MTADS data set suffered from a high number of false alarms due to anomalies with a geological origin (caused by the cart bouncing up and down). In addition, the operating point was very conservative and many non-UXO were excavated after recovery of the last UXO in the dig list. The results from the EM61 cart were also very good, although 24 FPs were required to excavate all 105 UXO (that weren’t in the can’t-analyze category). The lower data quality of the EM61 cart resulted in a larger number of can’t-analyze anomalies over metallic sources than the MTADS.
At Fort McClellan, all demonstration metrics related to discrimination of medium caliber projectiles were met. At the operating point, all but five of 119 targets of interest were recommended for excavation, with 34 false alarms. If the operating point was relaxed slightly, then all medium caliber projectiles would have been recovered with 51 false alarms. Retrospective analysis revealed that excellent discrimination performance could have been obtained by using a feature space comprising an early and late time feature extracted from the object’s primary polarization. Furthermore, the demonstrators found that these feature vectors could be approximated without fitting polarization tensor models to the data and by using just seven measurement locations around the template center. These approximate early and late time decay features were extracted from the sounding with the slowest decay (defined as the ratio of the 20th to 1st time channels).
Specific implementation issues related to the technologies demonstrated in this project include: