Classification Of Cued Metalmapper Data Using Data Mining Techniques

A total of 1444 MetalMapper cued soundings were inverted for polarizability transients and then analyzed using data mining techniques to classify anomalous responses as either likely targets of interest or unlikely to be targets of interest. Polarizability transients were parameterized using six scalar moments covering size, shape, and persistence. Additional parameters were generated using a curve fitting algorithm to calculate the best fit coefficients of a polynomial function matching the polarizability transients in log/log space. Parameters were then evaluated using principal component analysis and visual observation to reduce the parameter set to the most definitive elements. Data mining algorithms including the clustering approach K means, and the classifier multi-layer perceptron were used to evaluate training data with known anomaly sources. Approaches developed and refined using training data were then applied to responses with unknown anomaly sources to determine if they represented targets of interest or were unlikely to be targets of interest. Anomalies were ranked according to the likelihood that they represented a target of interest. All of the targets of interest were recovered within the first 13 percent of the ranked list.