Risky Removal: Developing a Holistic Understanding of the Risks of Redeveloping Sites Contaminated with Unexploded Ordnance

Abstract

In the United States, approximately 10 million acres of public land have been used for live-fire military exercises 1 and some fraction of the ordnance deployed here remains active and buried on site. As military facilities transition to civilian uses, unexploded ordinance (UXO) may endanger human lives or leach harmful chemicals into the environment. Thus, the United States government has invested considerable resources into finding and removing UXO1. Using current techniques, magnetometers and electromagnetic induction sensors collect geophysical data to identify anomalies (metallic objects that might be UXO) and use signal-inversion and feature-extraction algorithms to infer likely, but uncertain, locations and material properties. 2 Lacking clear decision guidance, risk-averse site managers often spend the majority of resources digging scrap metal with false-positive UXO signals, which is untenable given the vast scope of the problem and diverts scarce resources from higher-risk areas. Despite sensor advancement, remediation managers face a series of difficult and nontransparent decisions regarding risk, cost, mission, and sociopolitical trade-offs. These decisions are often dealt with ad hoc and fall short of risk-informed guidance regarding when to remediate, where to prioritize digging, suitability for future uses, and residual risks. Even with extensive digging, it is difficult to know that all UXO have been removed, making questions like “can you guarantee that this site is safe?” and, “how sure are you?” difficult to answer. Intuitively, we understand that risk involves more than sensor data and signal processing. Given two similar sites with similar distributions of identified anomalies, would a site manager be equally confident that remediation had ensured future safety if (A) one site is intended for a wildlife preserve and the other for public use; (B) one site has an oral history of live ordinance and the other of predominantly inert ordinance use; or (C) one site is near many excavated UXO and the other near excavated scrap with false-positives signals? In each case, advanced classification technologies may not differentiate between the contrasted examples, yet a rational person would expect risks to differ. Rather than indicating a failure on the part of sensor technologies, this highlights the importance of including nonsensor information. Bayesian Networks 3 (BNs) are particularly useful for transparently integrating diverse information to estimate probabilities. For example, suppose interviews, site histories, and geophysical information inform expert judgments about likely UXO concentrations. Experts define prior probability distributions (e.g., Beta distributions) for the proportion of UXO. Geophysical sensors provide additional information. Real-time excavation results update the distributions. Existing sensor analysis tools match the closeness between detected anomalies and object types. BNs can combine all these into predictive probabilities (Figure 1, left side). While BNs have occasionally been used in current classification efforts, they do not integrate dig results and rarely use data from nonsensor sources 2 .