Until recently, detection algorithms could not distinguish between buried unexploded ordnance (UXO) and clutter, leading to many false alarms. Over the last several years, modern geophysical techniques have been developed, merging more-sophisticated sensors, underlying physical models and statistical signal processing algorithms, with such approaches yielding reduced false alarms. For sites where anomalies are well separated, it has been shown that the combination of phenomenological models and advanced signal processing can markedly decrease the time required to remediate a site by classifying UXO and non-UXO items correctly. For highly contaminated regions, however, the signatures of multiple anomalies often overlap, vitiating the utility of many of the newer techniques. To address this problem, a synergistic use of advanced phenomenological-modeling and signal-processing algorithms would be beneficial.

The objectives of this project were to develop physics-based signal-processing approaches applicable to scenarios in which responses from multiple UXO and clutter items co-exist in a sensor signal, with the goal of discrimination; and use information-theoretic measures to define the types of scenarios for which UXO and clutter density is too high to reliably perform classification, necessitating a direct mechanical excavation of an entire region.

Both exact and approximate phenomenological models had been developed for electromagnetic induction (EMI) and magnetometer sensors, and these models had been used within the context of statistical signal processing algorithms. These models were used in this project to train the statistical algorithms, to develop clutter statistics, and to generate a database of simulated signals on which algorithms could be tested. These models were applied iteratively to the multiple, proximate object problem so that the effect of interaction terms could be considered rigorously. Both traditional and novel signal processing approaches were investigated to allow both the detection of the existence of multiple objects and the separation of the individual signatures from the cumulative signature. Specifically, the algorithms described in the blind source separation literature were utilized to extract the individual signals from the measured mixtures. Performance bounds also had been developed for detecting and discriminating UXO objects in isolation. Employing more sophisticated models, capable of handling the EMI signature of multiple UXO and anthropic-clutter items, extended these performance bounds to the case of multiple objects. Once the separation and discrimination algorithms were developed and tested on simulated data, they were tested on data measured in the field with both EMI and magnetometer sensors. Feature-based sensor fusion was also considered.

Simulations, test stand data, and field data indicate that independent components analysis (ICA)/ blind source separation (BSS) techniques can be used to extract individual ordnance signals from mixtures measured by EMI systems when UXO and clutter are in close proximity. The eigenvalue decomposition approach (EDA), which is less sensitive to correlated objects, provided the best performance of the approaches considered. While there are several caveats, and research remains to be done, this study demonstrated that ICA can restore classification performance from essentially chance levels when objects are closely spaced. Issues related to determining when there are multiple objects present appear most pressing, and may require more densely sampled data. Test stand results were promising, but a conclusive test on a realistic data set is necessary prior to any firm conclusions.

Phenomenological models allow quantification of the performance of UXO sensors as a function of object density. The models provide a realistic method of generating training data upon which algorithms can be tested. Such training data sets will hasten the transition from algorithm development to deployment, and also will facilitate assessment of algorithm robustness. The algorithms provide the ability to separate the signatures associated with different subsurface objects from a composite signature measured by a sensor. Accurate separation of such signatures will permit remediation of sites that cannot currently be considered using conventional techniques.