New technologies for detection and classification of buried unexploded ordnance (UXO) have the potential to significantly reduce the cost of munitions response projects. In particular, electromagnetic (EM) sensors developed specifically for this problem can reliably discriminate between ordnance and non-hazardous metallic clutter. The classification process involves fitting a physical model to observed sensor data and then using the parameters of this model to make inferences about the physical properties of a detected target.

ESTCP demonstration projects have shown that advanced classification with next generation EM sensors consistently outperforms commercial standard systems. To further the use of advanced classification in munitions response projects, decision support tools are required to help project geophysicists, managers, and stakeholders understand:

(1) How to best deploy available technologies for a particular remediation problem;

(2) How to ensure with high confidence that all targets of interest are identified following remediation efforts.

While significant advances have been made in the acquisition and processing of geophysical data for classification of buried munitions, the success of any classification strategy strongly depends on the site characteristics, including range of munitions types and clutter, geological background, topography, and vegetation. The objective of this project was to develop and validate the components of a decision support system (DSS) that will help individual site-managers and teams design surveys and data processing strategies to achieve optimal discrimination performance at the lowest attainable cost for a given site.

The technical approach to this project considered two major research topics:

(1) Performance prediction. Techniques were developed to model the performance of advanced sensors at a site. Both physical modelling of target response thresholds and developed statistical models were considered to assess the feasibility of classification under site-specific conditions.

(2) Risk assessment. Statistical models were developed to assess the posterior probability that targets of interest (TOI) remain in the ground following remediation efforts.

This research project completed work on the following topics:

(1) Optimizing detection with multistatic sensors. As monostatic sensors were replaced with multistatic, multi-component sensors for dynamic detection surveys, a threshold analysis tool was required to determine the minimum anomaly amplitude expected for a TOI at a specified maximum clearance depth. Analysis tools were described that were developed for threshold analysis with MetalMapper and TEMTADS2x2 sensors. An algorithm for objectively selecting time channels and receiver components for target picking with dynamic multistatic data was also developed. The approach defined a detection channel that was a linear combination of received channels. The weightings of received channels comprising the optimized detection channel were estimated by maximizing the expected signal to noise ratio for a target of interest at a specified maximum clearance depth. Finally, delineation of regions in dynamic detection data where classification cannot be applied was considered. It was shown that a singular value analysis of the received data can be used to filter out isolated anomalies and reduced the size of areas designated for “mag and dig" operations.

(2) Performance prediction. First, efficient Monte Carlo (MC) methods were developed for predicting the variability of estimated polarizabilities under site-specific conditions (e.g. sensor noise, target density, etc.). This approach provided a rigorous means to simulate the probability of correct classification of specified UXO and non-UXO.

Once EMI sensor data were collected and inverted, initial predictions of classification performance were updated with site-specific information. A number of data and model quality metrics to assess the overall difficulty of the classification task. These included: mean polarizability misfit with respect to library items, signal to noise ratio, and a metric that uses the point-to-point variability of soundings or polarizabilities to determine the number of channels that can be used for classification. These metrics were combined into a “Dataset Degree of Difficulty" (DDD) that categorized the classification difficulty at the site. This approach gave the data analyst an objective measure with which to assess the feasibility of classification at a site using available information. However, retrospective analysis of cued MetalMapper data sets indicated that the DDD was predictive of average classification performance and could not reliably predict false alarm rates at the point where all TOI were identified. This difficulty was addressed with a regularized linear regression algorithm that directly learned the relationship between performance metrics and the observed false alarm rate. Finally, the regression model was extended to generate predictions at an early stage of a project (prior to data collection) by imputing missing metrics with values drawn from past sites.

(3) Risk assessment. A random compliance sampling approach was suggested for UXO risk assessment, and this approach was extended to account for the bias in prioritized digging, thereby reducing the number of excavations required to test for outlying UXO. Then, methods for identification of outliers to the distribution of UXO via generative models of the receiver operating characteristic (ROC) were discussed and compared. Next, how seeded items emplaced for quality control can be used to increase confidence in the classification process was considered, and this process was modeled by constraining the ROC model. Finally, the problem of identifying novel, or unique, UXO with prioritized validation digs was briefly studied. A metric that combines features of the geophysical model estimated for each detected target to identify novel UXO was proposed. The metric required no prior information about the UXO present at a site.

The methods developed under this project will aid managers in designing a cost-effective remediation effort prior to deployment and in adjusting and optimizing survey design and data processing as more information becomes available. These tools will help users and stakeholders understand the potential benefits and limitations of advanced classification.