Recent research has led to the development of modern geophysical techniques that merge sophisticated sensors, underlying physical models, and statistical signal processing algorithms. These new approaches have dramatically reduced false alarm rates for unexploded ordnance (UXO) detection, primarily on data from sites with relatively benign topology and anomaly densities. Significant work remains to transition these advances to robust, field-ready technologies. To address the limitations associated with current sensors, SERDP and ESTCP have supported the development of a new generation of UXO sensors that produce multi-axis vector or gradiometer measurements, for which optimal processing has not yet been carefully considered or developed.

The objective of this project was to develop algorithms that provide UXO classification capability for multi-axis sensors that are superior to single-axis solutions.

Technical Approach

Previous SERDP-sponsored research was leveraged and extended to 1) develop exact and approximate phenomenological models for multi-axis sensors, 2), develop, and apply to field-collected data, statistical signal processing algorithms which maximize UXO classification performance using data measured by multi-axis sensors, and 3) develop active learning and kernel-based algorithms to enable algorithm training and application across test sites.


The modeling efforts focused on developing phenomenological models for the Berkeley Unexploded Ordnance Discriminator (BUD) sensor and a time-domain electromagnetic induction (TEM) system whose excitation waveform is on all the time (ALLTEM). Both of these sensors employ multiple transmit and receive coils at different orientations. Although the sensors were not fully modeled under this effort (i.e., some of the transmit and/or receive coils were not yet incorporated into the model), there is generally good agreement between the measured data and the modeled sensor response. In addition, the sensor models are currently being completed so that all transmit and receive coils will be fully incorporated into the models. As sensors continue to be introduced and refined, it will be important to continue modeling efforts for the new sensors so the benefits of model-based signal processing will continue to be realized. The statistical signal processing efforts focused on developing physics-based statistical signal processing approaches appropriate for data measured by multi-axis sensors, including model inversion, feature selection, classifier design, and sensor management, and evaluating the proposed approaches on field-measured BUD and ALLTEM sensor data. Results indicate that applying statistical signal processing algorithms to multi-axis sensor data has the capability to provide improved UXO classification performance. Results also showed that adapting standard algorithms, such as numerical least squares for model inversion, to the specific sensor and task at hand can improve their robustness. To this end, further work focused on continuing to customize these algorithms to ensure robust model inversion would be worthwhile. Similarly, other approaches which may improve robustness should be explored. Results obtained under the present effort have indicated that inversion performance may be improved by judiciously selecting the measurements to include in the model inversion, or by selecting in a principled manner one inversion from a group of several which are equivalent in terms of goodness-of-fit, and continuing to explore these ideas to improve model inversion performance and robustness could be beneficial. Continuing to improve robustness throughout the processing chain by extending the work on feature selection and classifier design would also be advantageous. For example, results obtained here indicate that feature selection performance may be sensitive to the performance metric that is optimized, so there may be opportunities to improve robustness by optimizing the performance metric utilized in feature selection.


The efforts related to active learning and kernel-based algorithms focused on developing approaches that would enable data collected across multiple sites to be integrated within the classifier. Results obtained here show that active and multi-task learning may provide frameworks in which the classifier parameters can be adapted within a data collection, or across data collections at multiple sites, to improve classifier training and UXO classification performance. Continuing to explore improvements to classifiers that would make them more robust to uncertainties inherent in the field would be valuable. For example, matching pursuits appears to hold promise for UXO classification in situations in which there are multiple spatially overlapping target signatures within a set of measurements. The models and algorithms developed under this program provide improved UXO classification for multi-axis sensor data. Further improvements may be possible by focusing on improving the robustness of model inversion, of the classifiers, and in the presence of uncertainties inherent in the field, such as overlapping signatures or varying clutter characteristics. An additional benefit of improving algorithm and processing robustness is that more of the processing may be automated, reducing the need to rely on human experts to analyze the data.