Until recently, detection algorithms have not reliably distinguished between buried unexploded ordnance (UXO) and clutter, leading to many false alarms. Over the last several years, modern geophysical techniques that significantly reduce false alarms while maintaining good detection have been developed by merging more sophisticated sensors, underlying physical models, statistical signal processing algorithms, and adaptive training techniques. Based on the characteristics of known target and non-target signals, a classifier algorithm is typically designed to determine the labels of the remaining unlabeled data, that is whether each unknown anomaly corresponds to ordnance or clutter.

Advanced signal processing algorithms traditionally require extensive training data. In this project, researchers have developed new active-learning algorithms that address the limited quantity of labeled data available for algorithm development and training. This is a ubiquitous problem in UXO sensing, of interest for airborne synthetic aperture radar (SAR), ground-based magnetometer, and electromagnetic induction (EMI) sensing data.

The objective of this project was to develop active learning and semi-supervised learning techniques that are applied to digital geophysical data sets. These techniques use algorithms that adaptively delineate regions that are clean of UXO. Additionally, for regions that are not clean, digital geophysics classification and detection algorithms detect UXO without requiring excavation of all items.

The active-learning algorithm developed under this project asks the following question: If it were possible to acquire a label for one of the unlabeled signatures under test, which would be most informative to classifier design? For SAR applications, the label may be acquired by employing personnel at localized positions on the ground, while for magnetometer and EMI sensors, the labels are acquired by excavation. The active-learning algorithm is executed by sequentially asking the above question, until the expected information from the next label no longer provides enough additional information to justify the cost of acquiring further labels. Therefore, the algorithm has a natural means of defining when to stop the active-learning process and make final classification decisions on the remaining unlabeled data (using a classifier based on the original labeled data and the new labeled data determined via active learning). Researchers have extended the active-learning formulation developed for missing labels to the general case of missing sensor data and/or missing labels.

This basic formulation has been successfully developed under this project and tested on geophysical data from munitions response sites in the Badlands, South Dakota, and Jefferson Proving Ground, Indiana. The technologies developed have transitioned to ESTCP project MR-200501, Demonstration and Validation of Advanced Digital Geophysics for Sensing UXO, for further development and demonstrations on additional data sets.

The U.S. Army Corps of Engineers Center in Huntsville reports that, on a typical job, about 75% of the remediation costs result from excavating non-ordnance targets. The techniques developed in this project achieve high UXO detection and significantly reduce total excavations, which lead to substantially reduced UXO remediation costs.