Unexploded ordnance (UXO) discrimination is achieved by extracting parameters from geophysical data that reflect characteristics of the target that generated the measured signal. Model-based parameters are estimated through data inversion, where the optimal parameters are those that produce acceptable agreement between observed and predicted data and satisfy any prior information for the target. These parameters then are used as inputs to statistical classification methods to determine the likelihood that the target is, or is not, a UXO. In recent years, various parametric modeling methods have been advocated for simulating electromagnetic induction (EMI) data from UXO and clutter. The optimal model minimizes a data misfit function (such as a least squares measure) and satisfies any prior information for the target. The choice of forward modeling method has a large impact on inversion, and therefore discrimination, performance. Ideally, a model can accurately reproduce the data with a minimum number of parameters, while being computationally efficient.
The objective of this project is to delineate the circumstances for which a particular type of modeling should be chosen and to generate methodologies and software that would allow the user to more efficiently extract meaningful parameters from data and thus improve UXO discrimination.
This project concerns the process of extracting physically realistic models of buried UXO from time- and frequency-domain electromagnetic data. Improved parameter estimation methodologies will be developed by defining strategies for choosing optimal forward models for data inversion and through quantifying and understanding the effect of data quality and model complexity on the recovered parameters. Researchers will determine the conditions for which charge-type model might yield superior information compared to dipole models. For cases where the dipole model is sufficient, researchers will determine which parameterizations are most useful for discrimination by examining the data quality required to constrain the two-transverse polarizations for UXO targets and three unique polarizations for non-axial symmetric targets. Strategies for determining the number of independent polarizations will be developed. For cases where the three polarizations are poorly constrained, researchers will determine if using two polarizations and relying on the misfit as the significant indicator for non-UXO targets is a better approach. To determine when a multi-peak anomaly should be processed as a single target or as a pair of targets, data features, such as separation of dipole peaks and differences in decay characteristics of the two peaks, first will be used. Statistical and information theoretic methods of determining the optimal model will be utilized. Practical diagnostic procedures will be developed to assess which data anomalies are of high enough quality to warrant inversion and which inversion procedure should be used. A “Figure of Merit” concept will be developed as a diagnostic for deciding whether inversion is relevant and what level of model complexity is appropriate for the inversion of time-domain electromagnetic data.
This research will provide tools that improve on the existing capabilities of reliably discriminating between hazardous UXO and nonhazardous metallic items, by improving the quality of parameters that are used as inputs to classification methods.