Objective

To classify UXO in survey data, an inductive learning method called ensembles will be used in conjunction with neural networks. The ensembles combine the output of many separate predictors.

Conducting unexploded ordnance (UXO) surveys using Global Positioning System (GPS)-guided geophysical mapping techniques is rapidly becoming the industry standard. Despite the investment in creating improved data, the use of physics-based signature analysis algorithms, and analyses employing both magnetometry and electromagnetic induction (EMI) sensors, clearing UXO ranges invariably requires digging 5-100 items for every recovered intact ordnance. In the recently completed SERDP SEED project MR-1354, the project team investigated several machine learning approaches to automate the analysis process and develop classification approaches based on the physical predictions of the dipole fitting routines used in analyzing magnetometry data. These included Artificial Neural Networks (ANN) and Genetic Ensemble Feature Selection (GEFS) approaches, trained on ground-truthed magnetometry survey data to analyze and classify blind data. In each case, the machine learning approaches gave inferior results to those of a human analyst interactively analyzing targets using dipole fitting routines and visual cues to make classification predictions. Researchers turned to resampling the raw data arrays, using a machine learning approach to select candidate targets, then analyzing those targets applying shape function information and shape filters resident in an existing commercial software product, Feature Analyst. Working with vehicular Multi-Sensor Towed Array Detection System (MTADS) survey data from a ground artillery range, this approach dramatically reduced the false alarm rate over that of the human analyst. This analysis, designed to reduce false positives, unfortunately also mildly increased the false negative rate.

The objective of this project is to develop a more sophisticated automated screening approach (i.e., target picker) that increases the inclusion of all viable UXO candidates, while continuing to reject non-UXO. Specific objectives include: (1) further evolve the characterization of size and shape parameters for use by the pattern recognition algorithms, (2) develop a probabilistic-based ranking system that can be adjusted to either maximize clutter rejection or to reduce false negatives, and (3) investigate the fusion of physics-based and size- and shape-based parameters.