The objective of this project was to demonstrate the ability of advanced analysis techniques, previously developed with SERDP support, to achieve high-fidelity discrimination between buried munitions or targets of interest (TOIs) and clutter. These advanced methods were applied to next-generation electromagnetic induction (EMI) sensor data acquired as part of the live site demonstrations. Application of the SIG statistical learning approach to UXO discrimination was demonstrated at three sites: Camp Butner, North Carolina; Pole Mountain Target and Maneuver Area, Wyoming; and Camp Beale, California. This technology was further developed and validated under previous ESTCP efforts by SIG and Duke University.

This project combined several advanced techniques incorporating and centered on a Bayesian approach to quantify statistical uncertainty—automated anomaly segmentation algorithms, physics-based target/sensor models, robust inversion algorithms for feature extraction, algorithms for data-driven feature selection, supervised and semi-supervised Bayesian classification algorithms, information-gain-based active learning for efficient label selection, multi-task learning (MTL) for sharing information from one site to the next, and principled cost- and cross-validation-based threshold selection. Each of these techniques plays a role in an overall discrimination-based approach to site remediation. The statistical rigor of these approaches enabled them to better inform decisions both in data collection (improving efficiency) and in target declaration (ensuring accuracy). The techniques represent the pairing of advanced data collection techniques and optimization of state-of-the-art digital geophysics.

The Camp Butner demonstration validated the robustness of key SIG technologies for sensor/target models, feature selection, classification, and active learning. The non-linear classifier outperformed the linear classifier. Both linear and non-linear classifiers would have left more than 75% of the clutter in the ground. The stopping point for both classifiers left UXO in the ground however. Two of these anomalies could have been captured earlier by selecting additional features.

The Pole Mountain demonstration tested MTL, where information from previous sites is incorporated into the current classification. The multi-task classifier outperformed the single-task classifier. All performance objectives were achieved. All of the UXO were classified as targets. The number of false alarms was fewer than 30% of the total. A generative model also was tested, and it reduced the number of false alarms to 32. This represents about 2% of the total false alarms.

The Camp Beale demonstration sought to validate and automate the SIG learning process using next-generation EMI sensor data for discriminating TOIs. This process includes three major components: feature extraction, site learning, and excavation. Results by sensor follow:

• TEMTADS 2x2, BealeTrees - All UXO were captured in the dig list with ~160 false alarms.
• BUD, BealeTrees - The two methods for selecting a single sounding at an anomaly, ‘symmetric’ and ‘closest’, captured all the UXO. The ‘closest’ method performed better than the ‘symmetric’ method in the number of unnecessary digs required to reach the last UXO: 180 vs. 200.
• CH2MHILL MetalMapper, BealeOpen - The final dig list missed two 37mm and one industry standard object (ISO). An additional 150 digs would have been required to capture all of these UXO.

Each of the classification approaches met all of the project objectives—maximize correct classification of UXO and non-UXO, specify a no-dig threshold, and minimize the number of anomalies that could not be analyzed. MTL required fewer training data for discrimination. Predictions based on the MTL model also had fewer false alarms than the single-task model. The generative model, however, outperformed both of the discriminative approaches in terms of number of false alarms. With the generative approach, all the UXO were revealed with only 32 unnecessary digs.

The demonstration results highlight the need to use different modeling approaches at different sites. Future work will focus on using generative and discriminative approaches synergistically based on adaptive estimates of site difficulty.

The software for the current SIG Isolate technology is based on MATLAB® and is not freely available. While the software is currently used by the experts who wrote the system, transitioning to minimally trained users is a goal of software development. Future demonstrations will be used to mature this software.