A favored approach to unexploded ordnance (UXO) clean up is based on data derived from computer-generated models, rather than raw physical data. Whereas previous approaches perform UXO declassification directly from sensor data, the model-based approach performs a joint transformation of data to derive physical parameters (position, depth, and orientation), allowing the classification to be based primarily on shape information and intrinsic variables.
This project focuses on the development of data fusion techniques for the best available existing sensor suites to better allow discrimination between intact ordnance and the typical clutter associated with target and bombing ranges. Specifically, the Naval Research Laboratory (NRL) intends to develop software techniques to allow the discrimination of intact ordnance from ordnance explosive wastes (OEW) using arrays of full-field magnetometers and time-domain electromagnetic sensors as the primary detection tools.
Based upon initial studies with the Mobile Towed Array Detection System (MTADS) data from Twentynine Palms, California, and the Badlands Bombing Range, South Dakota, the researchers expect that by using this data-based approach, 70-90 percent of targets can be selected and analyzed automatically. Following the automatic target selection process, model-based quantitative magnetic and electromagnetic (EM) routines will be used to solve the inverse problem for target position, depth, shape, and orientation. Then a probabilistic classifier (Bayesian or neural net) will model the output to identify likely UXO type and distinguish UXO from ordnance explosive waste or other clutter. Finally, an analyst, as a backup to the automatic target picker, will work interactively with the individual graphical images to pick targets that are not common to the magnetometer and EM data sets (or for which the automatic target picker solution was not accepted) for subsequent analysis by the physics-based target fitting routines.
When integrated into an operational unexploded ordnance survey system such as MTADS, this data analysis system will reduce target analysis time by up to 50 percent. Location information, including position, size, and depth, also is expected to be mildly improved. The major benefit of this analysis will be a significant improvement in the ability to differentiate UXO from OEW and other clutter.
Successful fusion of seven data sets from the Badlands and Blossom Point sites were demonstrated using PCA analysis. The first principal component describes the fused data sets when covariance or correlation matrices with standardization were used. In addition, digital filtering was investigated. Two linear filter functions, Gaussian and Laplacian of Gaussian (LoG), have been investigated. For both types of filters, various widths and window sizes were studied to characterize the performance of the filters on MTADS images. Features were enhanced using filtering techniques. It was determined that PCA analysis following digital filtering provides sensor fusion and enhanced images.