Learning Machine Identification of Ferromagnetic UXO Using Magnetometry

The fundamental problem in applying geophysical mapping to locate unexploded ordnance (UXO) is distinguishing true UXO from non-UXO. Enhancing the accuracy of UXO detection has multiple benefits, especially in the areas of cost savings and safety. We investigated discrimination approaches using both magnetic field data and numerically modeled data. Libraries of total field magnetic (TFM) responses were calculated using finite element modeling for three UXO types found at a Montana National Guard training site. UXO model parameters were varied over ranges of azimuth, declination, and depth resulting in approximately 600 models per UXO type. The modeled responses of finite-element model (FEM) and actual TFM field data were then used as training data in discrimination and classification approaches comparing neural networks (NN), random forests (RF), and support vector machines (SVMs). The prediction targets in the training process comprised three classes: 1) binary [UXO or noninteresting object (NIO)]; 2) multiclass (UXO round type and NIO); and 3) classes derived from multiclass self-organizing feature map (SOFM) analysis. The multiclass SOFM targets generated from site-specific field data were found to be optimal for UXO discrimination. The best performing combination of class selection types using recentered data for UXO detection rates of 100% resulted in a false alarm rate (FAR) of 28%.