This research program was focused on advanced technologies for detection and discrimination of military munitions. The underlying premise of the program was that there is an inherent limitation in the information content associated with magnetometer and electromagnetic induction (EMI) sensors deployed for unexploded ordnance (UXO) cleanup. To optimize UXO classification, one must integrate all available information, both within the measured data itself and within a priori knowledge one may possess.

An important class of prior knowledge is represented by the sensor physics, and by placing as much physics as possible into the models and classification features, one removes the need to rely on the limited sensor data to infer such phenomenology. Statistical classifiers are also required to maximize the information extracted from the measured data to infer the unknown model parameters. Further, the statistical classifiers may be used to appropriately exploit other forms of information inherent to the data. For example, while performing classification one may exploit the contextual information provided by all of the unlabeled data at a given site, while also appropriately leveraging related information in data measured at previous  sites.

The overall objective of the research was to integrate advanced Bayesian statistical models and classifiers with leading geophysical models to enhance the ability to extract information from limited sensor data, with the goal of markedly improving UXO classification performance on complex cleanup missions.

Technical Approach

The technology was directed toward general magnetometer and EMI sensors. A key aspect of the research was to develop sophisticated but practical technology appropriate for real-world UXO cleanup. The technology was also directed toward difficult geology, terrain, and complex ordnance and clutter distributions.


This research program placed a focus on integrating the statistical inference engines developed at Duke University with the sophisticated physics-based models developed at the University of British Columbia (UBC)/Sky Research (Sky). The particular statistical techniques into which the advanced geophysical models were integrated include semi-supervised learning, multi-task and life-long learning, and active learning. This project also developed new techniques that explicitly account for the imbalance in UXO and non-UXO items at a typical site, which is of significant importance when computing the risk associated with leaving an item unexcavated.


By integrating the Duke and UBC/Sky technology, the Bayesian statistical models have been aided by improved geophysical models, and vice versa. This new technology has the potential to significantly improve the Department of Defense’s ability to conduct practical UXO cleanup. The experience of the investigators within the ESTCP demonstration studies has guided selection of the open research questions to be investigated, advancing the likelihood that the research products will constitute new science while also being of importance to practical UXO cleanup.