In 2003, the Defense Science Board observed: “The problem is that instruments that can detect the buried unexploded ordnance (UXOs) also detect numerous scrap metal objects and other artifacts, which leads to an enormous amount of expensive digging. Typically 100 holes may be dug before a real UXO is unearthed! The Task Force assessment is that much of this wasteful digging can be eliminated by the use of more advanced technology instruments that exploit modern digital processing and advanced multi-mode sensors to achieve an improved level of discrimination of scrap from UXOs.” Significant progress has been made in discrimination technology. 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 of an entire decision making process.
The objective of this project was to apply and evaluate classification algorithms to demonstrate that some non-UXO items can be classified correctly and hence left in the ground, while maintaining a given level of detection performance.
The classification algorithms were demonstrated on three sites. They were initially validated at Fort Ord Seaside, California and then applied as part of the ESTCP Classification Program at the Former Camp Sibert, Alabama and the Former Camp San Luis Obispo (SLO), California. Results from early implementations show that the demonstrators were too conservative in the application of their methods. The demonstrators did not classify a large number of targets because they labeled them as “can’t analyze”, which can be due to factors such as signal-to-noise ratio (SNR) and data quality. Targets identified as “can’t analyze” are treated as potential targets of interest and, in large quantities, can significantly reduce the potential for cost savings at a site.
Building on lessons learned from the Ft. Ord Seaside and Camp Sibert demonstrations, the demonstrators refined their methods and the active learning algorithms produced outstanding results at Camp SLO using MetalMapper sensor data. The demonstrators were able to analyze a much larger percentage of site targets. The majority of targets of interest and non-targets of interest were readily recognized and classified as such with high confidence.
The Department of Defense will benefit from the reduced number of total excavations required to clean a given site. The adaptive algorithms demonstrated provide a mathematically rigorous means of adaptively augmenting a training data set, based on the site-dependent observed data. This procedure provides an ordered dig list. Moreover, the graph-based semi-supervised algorithm provides a means of performing classification based on all observed data at a given site, not one feature vector at a time. Both of these techniques have resulted in significantly reduced total excavations, while achieving high UXO detection. The techniques demonstrated in this project represent the state-of-the-art in digital geophysics, and their success has the potential to transform UXO cleanup methods.