The objective of this project was to advance classification capabilities of the production community by enhancing and streamlining UX-Analyze’s workflow and functionality; by offering training documents, datasets, and workshops; by baselining classification performances through classification demonstrations; and by collaborating with multiple production firms, each of whom hold key geophysical services contracts for the government, as they perform data analysis services in support of future classification demonstrations.

Technology Description

The technology and analysis algorithms to be demonstrated were developed in ESTCP projects MR-200810 and MR-200910. By virtue of the past projects, UX-Analyze has been fully integrated into Oasis montaj as a menu-driven set of functions for geophysical target classification, modeling, and analysis. These functions permit users to effectively discriminate munitions targets. At its core, UX-Analyze provides four fundamental capabilities. These relate to 1) visualizing the measured data and anomaly extracting (not needed for cued data collections); 2) performing data inversions; 3) deriving a classification decision metric; and 4) documenting the decision. UX-Analyze was designed to automate analysis tasks that do not require user expertise, quantify the decision process, and produce transparent decisions. In collaboration with built-in Oasis montaj functionality and capabilities, UX-Analyze provides the tools, procedures, and graphics necessary to characterize, classify, and rank order source objects based on their electromagnetic induction (EMI) signature when interrogated by an advanced EMI sensor. Among other things, UX-Analyze inverts measured data for the targets’ intrinsic magnetic polarizabilities, which, in turn, provide information regarding the targets’ size, shape, wall thickness, and material type. It also classifies anomalies based on how similar their polarizabilities are to unexploded ordnance (UXO) signatures. Although this process sounds complex, perhaps because of the data inversion and classification step, UX-Analyze makes it quite simple to implement.

Demonstration Results

NAEVA Geophysics, Inc. (NAEVA) processors correctly classified all targets of interest (TOI) and achieved a substantial reduction in the number of required digs. The limited ground truth available restricts the extent of retrospective analysis possible. Four of the targets categorized as “High Confidence TOI” in ranked list Ground Truth Evaluation 1 have ground truth indicating that cultural debris was found at these locations, but no recovered depths or item descriptions were provided. Removing these items from the scoring results produces a moderate improvement in classification.

Thresholds were assigned for several parameters as part of the classification process. Threshold values were applied using the “Set thresholds and prioritize” tool contained in the “Classify and Rank” routine. The TOI threshold was determined based on a substantial drop-off in the visual similarity between the extracted polarizabilities and library TOI, training data ground truth, and intrusive results from the top ranked targets on the dig list.

Implementation Issues

By correctly identifying high confidence clutter objects based on their EMI signatures prior to excavation, savings can be realized by leaving the clutter in the ground or by changing the manner in which the excavations are performed. The estimated cost to clean up UXO on known Department of Defense land without classification exceeds $14 billion. Results of past ESTCP-led live site classification demonstrations at seven sites, however, show that on most sites, new technologies can distinguish metallic scrap 70-90% of the time.

A number of anomalies were selected from the ranked list to better characterize on-site TOI and refine classification thresholds, and from representative samples from identified unique clusters to assist in refinement of the generic library to a site specific library. Unnecessary entries from the generic library were excluded and any additional TOI or unique clutter identified from the training data request were incorporated.