Cleanup of subsurface unexploded ordnance (UXO) at military installations and training ranges is an expensive and time-consuming challenge. While the goals in UXO remediation are very clear, to clean up all UXO as efficiently as possible, little effort has focused on designing robust and efficient signal processing strategies with the specific performance goal dictated by the regulators in mind. This project originated from the hypothesis that performance and robustness may be improved over the classical approaches by specifically considering the desired operating point of the UXO discrimination strategy (100% detection) during the construction of each stage of the signal processing sequence that is needed to make the “dig/no dig” decision. From a statistical decision theory perspective, operating at this specific point has implications that may impose a strong preference for certain processing techniques in the UXO/clutter discrimination process. The objective of this project was to conduct a preliminary investigation into the potential benefit of awareness of the specific performance criterion (100% UXO detection) in each stage of the UXO discrimination processing strategy. This work should lead to new strategies for training and classification and may suggest guidelines for all stages of data processing.
This project consisted of several large-scale classification studies to carefully analyze the performance of different classification algorithms and the effects of training data when operating at 100% UXO detection. The data used in this study were collected during the Camp San Luis Obispo, California, demonstration with the MetalMapper time-domain electromagnetic sensor. The various classification algorithms included in this study provide a diverse representation of the different theoretical approaches to pattern classification and allow for comparison of the effect of different classifier properties on performance at the 100% detection operating point.
Across a large number of experiments, strong performance was consistently observed with a nonparametric classification algorithm that makes decisions locally in feature space based on neighboring training samples. Such a classifier shares properties with the library-matching classifiers that are often used in the UXO research community for classification based on the polarizability curves. Additionally, preliminary analysis of a method for evaluating the outputs of the model inversion procedure shows potential for identifying potential outliers (which drive performance at the 100% detection operating point) and merits more careful follow-on analysis.
This project provides evidence that the desire to operate at 100% detection may lead to a preference for certain algorithms in the different stages of the discrimination strategy. Careful consideration and selection of methods used in each stage of the discrimination strategy may greatly impact performance at the 100% detection goal. This preliminary work towards developing guidelines for classifier design, use of training data, model inversion, and feature selection in the UXO discrimination algorithm will eventually lead to more robust methods of data processing.