Over the past few years, sophisticated passive and active metal detectors have been coupled with state-of-the-art (SOA) global positioning systems (GPS) to provide increasingly accurate real-time localization capability when conducting unexploded ordnance (UXO) surveys. However, clearing UXO ranges with these automated techniques still requires digging 5-100 items for every recovered intact ordnance item. In an attempt to improve discrimination and reduce false alarms, target analysis techniques have been developed to process UXO survey data sets, emphasizing statistical analysis approaches applied to the output parameters of physics-based target fitting algorithms.

The objective of this project was to go beyond the use of physics-based parameters when making decisions about ordnance classification. The project team incorporated shape representation and similarity based on anomalies in the mapped data files to extract and exploit image features related to the target signatures. This provides an additional input for a discrimination/classification decision.

Shape information is an important component of the semantic content of the UXO target image and is a primary component of the visual decision making process used by the human analyst in the current interactive data analysis approach. Because this information is important to the human in-the-loop, if it can be quantified and incorporated into the machine analysis of the data, it will provide an important classification tool. Indeed, if the target analysis process is ever to be fully and effectively automated, implementation of this step is imperative.

Technical Approach

This project employed existing data sets taken by the Vehicular Multi-sensor Towed Array Detection System (MTADS) at the Badlands Bombing Range (BBR), South Dakota during 2001. In preparation for that survey, a 10-acre area near the bull’s eye (which had previously been cleared of UXO in 1999) was seeded by the Engineer Research and Development Center (ERDC) with 25 projectiles (ten 105-mm projectiles, ten 155-mm projectiles, and five 8-in projectiles). These inert projectiles were buried in a variety of orientations at depths up to the maximum expected self-burial depths for these ground-fired projectiles. One additional pre-existing high explosive (HE)-filled buried projectile was discovered in the 10-acre area. This 10-acre area and an additional 100 acres were surveyed by the Vehicular MTADS. The entire data set was interactively analyzed using standard techniques by an analyst using the MTADS data analysis system (DAS); a target list was prepared and all targets on the list were dug by UXO technicians.

The pattern recognition team on this project was provided the complete ground truth (locations and identities of all dug targets, including ordnance and non-ordnance [metallic scrap] items from the 10-acre seeded area). The data for the additional 100 acres was provided to the pattern recognition team as a mapped data file. The dig results and target location ground truth for this area were blind to the group until final analysis results had been submitted to the principal investigator.

To begin the study, the pattern recognition team applied statistical analysis approaches based on the physics-derived fitting parameters resulting from the previously completed interactive analysis of the seed target area. These techniques are similar to those that have been applied by others using data sets from other ranges. They had not been previously applied to data from the BBR.

The culminating step in this research project involved (a) developing parameter sets that describe shape information from the magnetic anomaly image map and (b) developing techniques to automatically filter these parameters and learn from them by submitting them to an inductive learning algorithm.


The following conclusions were drawn from this study:

  1. A human analyst, working with a physics-based modeling analysis system and high resolution mapped data files, can achieve excellent detection results at relatively uncomplicated sites like the BBR. However, the analyst’s ability to distinguish between intact ordnance and clutter still requires digging 5-20 targets for each recovered UXO.
  2. This project demonstrated that the physics parameters generated by a human-in-the-loop analysis approach, such as at the BBR survey, cannot be significantly improved by the use of inductive learning algorithms that focus on the physics-based fitting parameters. In other situations, researchers working with different data sets from more complex sites have drawn different conclusions. The generality of these observations may depend upon the complexity of the survey site, the quality of the data sets and the human-assisted analysis.
  3. In the early parts of this study, the use of raw spatial pixel values from the MTADS data, when appended in a cooperative analysis with the physics-based predictions, did not improve classification accuracy.
  4. In the final study, the first-pass technique that was developed provided a quick and effective automated target picker. With these data sets, this approach can effectively replace the human in-the-loop for target selection, or at a minimum, can provide a valuable aid to the analyst. This automated technique can be adapted to different fields and situations.
  5. The second-pass technique provided the ability to use shape and intensity information from the MTADS data to improve the automated discrimination of UXO and other metallic scrap. As a result, the new techniques hold great promise in reducing the numbers of non-UXO that must be dug.


In general, the results of the cooperative analysis of the BBR data using shape filters as a classification tool suggest some tantalizing possible improvements in classification. Using these pattern recognition tools in conjunction with either a primitive automatic target picker, or the physics-based analysis as a pre-screener, reduced the dig list by more than a factor of two. This significant advance in correctly identifying clutter targets came at the expense of some missed UXO targets.

Because of the limited resources available for this study, the scope of this project’s efforts had to be restricted. There are several areas where significant improvements and evaluations could be made. There was neither time nor resources available to implement a probabilistic evaluator in the shape function analysis. To critically evaluate the performance of these new tools, it will be necessary to develop this capability. It will allow the generation of receiver operating characteristic (ROC) curves and the evaluation of this performance approach against others. Equally importantly, it would allow classification thresholds to be adjusted to emphasize either UXO detection or clutter rejection.

Extensive ground-truthed data sets are necessary to achieve adequate training. It is important that the training sets have extensive UXO and clutter targets. Moreover, it is important that these training data be associated with the actual site to be analyzed. Indigenous clutter and geological interferences are critical contributors to noise. Noise sources ultimately establish the detection and classification floor because of signal-to-noise limitations.

The BBR site was chosen for this study because of its simplicity. From an ordnance perspective, it is almost a single use site. Background clutter (except near the bull’s eye) and geological interferences were low, and generally, UXO targets were well separated and isolated from each other. Finally, this is the most extensive area the project team was aware of that has been studied with high quality surveys and 100% remediated.

These shape function filters combined with a cooperative analysis approach need to be further developed and evaluated at other ranges against a more complex mix of UXO and background challenges.