Signal Processing and Modeling for UXO Detection and Discrimination in Highly Contaminated Sites

摘要

In this effort, we considered the problem of classifying closely spaced UXO. When two UXO are in close proximity, their signatures as measured by electromagnetic induction (EMI) sensors co-mingle, and traditional classification algorithms cannot be utilized. We explored the use of independent components analysis (ICA), a technique from the blind source separation (BSS) literature as a pre-processing step by which to separate the individual UXO signatures from the mixtures measured by the EMI sensor. Simulations indicated that this procedure could succeed and restore some level of classification performance in the case of overlapping signatures. Test-stand data also indicated that the ICA-based classification approach showed promise. Finally, testing the algorithms on data collected by NRL/WES suggested that in some cases ICA could be used as a preprocessing step and that closely-spaced UXO could be classified. During the course of this effort, several studies were performed to try to assess the limitations of the ICA-based approaches for discrimination of closely spaced objects. Correlation between object signatures can degrade performance, and alternate techniques based on ICA were considered as a remedy to this issue. Other issues including sampling density, noise susceptibility, and others were also investigated.