The objective of this project was to develop and validate innovative, robust, and practical approaches for unexploded ordnance (UXO) localization and classification under realistic (noisy, cluttered background) field conditions by combining advanced electromagnetic induction (EMI) forward and statistical signal processing methodologies. In a real field, the electromagnetic signals become convoluted with noise due to the instrument, magnetic soil and widespread background clutter. Therefore, the rationale for a statistical approach is to use an advanced statistical approach to reduce the impact of these noises to minimum.
This project provides mathematical fundamentals, physical meanings and practical realizations of forward, inverse and statistical signal processing approaches for UXO detection and discrimination at live UXO sites. The project team first developed and implemented advanced, physically complete forward EMI models, such as the normalized surface magnetic source (charge/dipole) model (NSMS), and ortho-normalized volume magnetic source (ONVMS) technique for accurately representing the EMI responses of subsurface metallic targets. The project team then combined these advanced models with EMI data inversion approaches, such as the gradient search and direct search-differential evolution, for advanced EMI sensor data inversion. Third, the project team extended the advanced statistical signal processing approaches, such as support vector machines and Gaussian mixture models, for discriminating UXO targets from non-hazardous anomalies. Finally, the combined advanced EMI forward and statistical models were applied to ESTCP live site UXO data sets.
Studies showed the excellent discrimination performance of the advanced models when applied to next-generation sensor data collected at various live sites, such as Camp Sibert, Alabama, San Luis Obispo (SLO), California, and Camp Butner, North Carolina as well as Aberdeen Proving Ground, Maryland test sites. The technology was able to single out UXO ranging in caliber from 25 mm up to 155 mm. In addition, the ONVMS technique provided excellent classification in both single- and multiple-target scenarios when combined with advanced multi-axis/transmitter/receiver sensor data.
The results clearly demonstrated that the suite of advanced modeling and classification tools developed by this project are robust and noise-tolerant and provide excellent classification results using real-world data collected by next-generation EMI sensors. ONVMS proved superior to NSMS and simple dipole model for inversion and classification purposes and shall remain the preferred method of analysis. The ONVMS-differential evolution- joint diagonalization (DE-JD) combination, supplemented by the classification algorithms, was further tested under ESTCP project MR-201101 using MetalMapper, Man-Portable Vector (MPV) handheld sensor, and 2 x 2 3D TEMTADS data collected at Camp Beale in California. Not only were the advanced EMI models able to classify all “easy seed UXO items”, they also managed to identify all other targets, no matter how unexpected or site-specific and as small as 3-cm fuzes.