Unexploded ordnance (UXO) cleanup is the number one priority Army installation remediation/restoration requirement The problem is enormous in scope, with millions of acres and hundreds of sites potentially contaminated. Before the UXO can be recovered and destroyed, it must be located. UXO location requires surface geophysical surveys. The geophysical anomalies caused by the UXO must be detected, discriminated from geophysical anomalies caused by other sources, and ideally identified or classified. Recent UXO technology demonstrations, live site demonstrations, and practical UXO surveys for site cleanup confirm that most UXO anomalies can be detected (with probabilities of detection of 90 percent or better), however there is little evidence of discrimination capability (i.e., the false alarm rates are high), and there is no identification capability. Approaches to simultaneously increase probability of detection and decrease false alarm rate and ultimately to give identification/classification capability involve rational multisensor data integration for discrimination and advanced development of new and emerging technology for enhanced discrimination and identification. The goal of multisensor data integration is to achieve true joint inversion of data to a best-fitting model using realistic physics-based models that replicate UXO geometries and physical properties of the UXO and surrounding geologic materials. Data management, analysis, and display procedures for multisensor data are investigated. The role of empirical, quasi-empirical, and analytical modeling for UXO geophysical signature prediction are reviewed and contrasted with approaches that require large signature databases (e.g., expert systems, neural nets, signature database comparison) for training or best-fit comparison. A magnetic modeling capability is developed, validated, and documented that uses a prolate spheroid model of UXO.

Unexploded ordnance (UXO) cleanup is the number one priority Army installation remediation restoration requirement. The problem is enormous in scope, with millions of acres and hundreds of sites potentially contaminated. Before the UXO can be recovered and destroyed, it must be located. UXO location requires surface geopbysical surveys. The geophysical anomalies caused by the UXO must be detected, discriminated from geophysical anomalies caused by other sources, and ideally identified or classified. Recent UXO technology demonstrations, live site demonstrations, and practical UXO surveys for site cleanup confirm that most UXO anomalies can be detected (with probabilities of detection of 90 percent or better), however there is little evidence of discrimination capability (i.e., the false alarm rates are high), and there is no identification capability. Approaches to simultaneously increase probability of detection and decrease false alarm rate and ultimately to give identification/classification capability involve rational multisensor data integration for discrimination and advanced development of new and emerging technology for enhanced discrimination and identification. The goal of multisensor data integration is to achieve true joint inversion of data to a best-fitting model using realistic physics-based models that replicate UXO geometries and physical properties of the UXO and surrounding geologic materials. Data management, analysis, and display procedures for multisensor data are investigated. A magnetic modeling capability is developed, validated, and documented that uses a prolate spheroid model of UXO. The electromagnetic modeling of UXO signatures is more problematic, and an intermediate quasi-empirical modeling capability (a simple analytical model modified to reflect measured signature observations) is explored.

(Cont.) Therefore, these coefficients readily lend themselves for use as features by which objects can be classified as likely to be UXO or unlikely to be UXO. To do such classification, the relationship between these coefficients and the physical properties of UXO and clutter, such as differences in size or body-of-revolution properties or material heterogeneity properties, must be found. This thesis shows that such relationships are complex and require the use of the automated pattern recognition capability of machine learning. Two machine learning algorithms, Support Vector Machines and Neural Networks, are used to identify whether objects are likely to be UXO. Furthermore, the effects of small diffuse clutter fragments and uncertainty about the target position are investigated. This discrimination procedure is applied on both synthetic data from models and measurements of UXO and clutter. It is found that good discrimination is possible for up to 20 dB SNR. But the discrimination is sensitive to inaccurate estimations of a target's depth. It is found that the accuracy must be within a 10 cm deviation of an object's true depth. The general conclusion forwarded by this work is that while increasingly accurate discrimination capabilities can be produced through more detailed forward modeling and application of robust optimization and learning algorithms, the presence of noise and clutter is still of great concern. Minimization or filtering of such noise is necessary before field deployable discrimination techniques can be realized.

The clearing of unexploded ordnance (UXO) is a deadly and time consuming process. The U.S. Government is currently spending millions of dollars to remove UXO's from bases that are closing around the world. Existing methods for detecting UXO's only inform the clearing team that a piece of metal is present, rather than the type of metal, either UXO, shrapnel, or garbage. A lot of time and money is spent digging up every piece of metal detected. This thesis presents the use of artificial neural networks to determine the type of UXO that is detected. A multi layered feed forward neural network using the back propagation training algorithm was developed using the language Lisp. The network was trained to recognize five pieces of ammunition. Results from the research show that four out of five pieces of ammunition from the test set were identified with an accuracy of .99 out of 1.0. The network also correctly identified that a tin can was not one of the five pieces of ammunition.

Unexploded ordnance (UXO) is a worldwide problem, which causes deaths and injuries to people living in post-conflict areas. It also prevents former military training sites from being returned to civilian use before they are properly cleaned. UXO cleaning is both dangerous and expensive. Metal detectors are the most common method used in UXO detection since reliable explosive detecting methods are not available at the current stage. Traditional metal detectors are unable to distinguish between targets of interest (TOI) and non-harzardous metallic clutter. As a result, UXO clean up could end up with about 95% of the digging to be clutter. To overcome this problem, new generation electromagnetic induction systems have recently been developed for subsurface target detection and classification. One of such system is the Time-Domain Electromagnetic Multisensor Towed Array Detection System (TEMTADS), which provides high fidelity EM data for target classification. UXO classification consists of three main steps: 1: Data collection; 2: Data inversion and target parameter extraction; 3: Target discrimination. In this thesis we present TEMTADS data sets inversion and processing approaches for cued (static) and dynamic (moving) measurements. Instead of using traditional matrix inversion and iterative search methods to tackle the highly non-linear problem of target locating, we employ the multiple signal classification (MUSIC) algorithm for fast and accurate estimation of target locations. The MUSIC algorithm is based on the orthogonality between the signal and noise subspaces in the multi static response (MSR) matrix of targets. In general, to identify the boundary between the two subspaces for actual data is a difficult task. To overcome this, joint-diagonalization (JD) is integrated into the processing. Namely, we use JD to estimate the number of sources presenting in a data set and to improve signal-to-noise ratio (SNR). The entire process is automated. Studies are done for test stand and blind data sets. Our results showed that the combined JD-MUSIC algorithm can be used to estimate target locations in near real time. The JD algorithm and Orthonormalized Volume Magnetic Source (ONVMS) model is extended for dynamic TEMTADS anomaly mapping and target picking. The algorithms were tested on data sets collected at live UXO sites in Camp Hale, Colorado. The study shows that JD and ONVMS provide underground target picking and preliminary classification capabilities, if clear background data can be provided. Comparison of inverted parameters between cued and dynamic data suggests that dynamic data can provide generally same value as cued data for classification purpose under good SNR conditions, thus the number of cued measurements needed can be reduced.

This report records work performed to develop data fusion algorithms to detect buried unexploded ordnance with magnetometers and ground penetrating radar. Task 1 characterized the performance of a cesium magnetometer, a gradiometer, and a 3-axis fiber optic magnetometer. Data which were measured and/or generated by validated models were used to develop a set of appropriate target features that could be used to identify and characterize buried ordnance. The applicability of applying techniques such as the use of neural nets, fuzzy logic and wavelets to the ordnance detection and identification problem were also evaluated.