For mobile, landscape view is recommended.
The objective of this project was to study discrimination capabilities of feature based characterization and classification techniques using standard survey data acquired by others at the unexploded ordnance (UXO) Standardized Test Sites in Aberdeen Proving Ground (APG), MD and Yuma Proving Ground (YPG), AZ.
The fundamental issues investigated in this project included the model used during characterization and the impact that classifier selection has on classification performance. After re-leveling and lagging the EM61 cart data, anomaly data for each data type were inverted using dipole, ellipsoidal, empirical, loop fit, joint frequency-time domain, and singularity expansion models. The resulting feature vectors were then classified with support vector machine (SVM), relevance vector machine (RVM), generalized likelihood ratio test (GLRT), and K-nearest neighbor (KNN) statistical classifiers.
Classification performance was evaluated using two metrics derived from receiver operating characteristic (ROC) curves; namely, (i) the total area under the curve and (ii) the probability of false alarms at 0.95 probability of detection. Five data sets were selected to include in this study based on data quality, type, signal-to-noise, and availability at appropriate intermediate processing stages. The datasets included time-domain EM61 (man-towed single sensor cart and vehicle-towed array), time-domain EM63, frequency-domain GEM-3, and magnetic data.
None of the classifiers or sensor/model combinations performed extremely well when the targets of interest (TOI) included 20mm- through 155mm-projectiles. Classification performance measures, defined here using area under the curve, were 0.8 for the best case(s). Additionally, all classifiers or sensor/model combinations produced multiple false negatives. False alarm rates, at a detection performance of 0.95, were as high as 0.95. Simplifying the problem by artificially limiting the UXO by size or by analyzing data acquired in a cued deployment did improve classification performances. Segmenting the UXO by size classes improved classification in direct proportion to the extent that the features of the UXO and clutter classes were separable. The cued data collection and comparison exercise showed improved classification capabilities with regard to deriving meaningful shape parameters when compared to data acquired on dynamic platforms.
The implication for future research is that new sensor technologies are required to realize acceptable classification performances. Given that there are multiple advanced sensors currently in development that are designed to maximize discrimination capabilities and mitigate limitations associated with single monostatic transmit-receive EMI systems, focusing future analyses efforts on their data is recommended.