The remediation of sites contaminated with unexploded ordnance (UXO) remains an area of intense focus for the Department of Defense. Current estimates place the total area of possibly UXO-contaminated sites at 10 million acres, with an overall cost of remediation with current methods and sensing technologies in the tens of billions of dollars. Fortunately, studies have estimated that up to 80% of typical sites of potential contamination are actually UXO-free. What is needed to take advantage of this ratio is a means to quickly and reliably scan large sites (on the order of 10,000 acres) in order to rapidly identify regions that are free of UXO and regions that must be subjected to more detailed and time-intensive examination and remediation with established UXO detection tools. Recent investigations have focused on wide-area assessments (WAA) aimed at rapidly determining the approximate density and spatial distribution of UXO objects over regions of wide area, rather than identification of individual UXO objects. Several WAA projects have been completed under the auspices of Strategic Environmental Research and Development Program (SERDP) and Environmental Security Technology Certification Program (ESTCP). These projects utilized various detection techniques, each with different strengths and weaknesses. However, no single sensing technology has been proven superior in WAA of UXO. It is therefore logical to examine data fusion approaches, which take advantage of all the available evidence, combining the strengths of each sensing technology while minimizing the weaknesses.
The objective of this work was to develop a data fusion framework that will form the basis of a cohesive data management and decision making utility for processing information acquired in the course of performing wide-area surveys of potential UXO remediation sites. This framework will be capable of capturing UXO-related information from all available data and effectively combining this information to provide site-wide assessments of the likelihood of UXO contamination that are more accurate than any single information source on its own. The final data fusion framework is intended to allow site managers to more efficiently direct the expenditure of time, labor and resources in remediation efforts.
The first year of this project examined potential WAA information streams, and determined the feasibility of feature selection methods for data fusion. The second year of the project focused on the development of an architecture and data fusion algorithms for the framework. Future work will focus on optimization and refinement of the methodology and algorithms with the development of a prototype implementation of the data fusion framework and its evaluation at several sites.
The major accomplishments of this project were:
The principal accomplishment this project was the development of a prototype data fusion framework suitable for WAA of UXO contamination. A key enabling technology was the development of a generalized method for processing input data feature streams from UXO WAA survey efforts. The development of this method is significant, as it requires only a limited number of specifications to be imposed on input features, allowing a wide range of feature sets and relationships to be formatted and input for data fusion. Such flexibility is crucial, as the disparate nature of the data and features available from potential WAA survey techniques presents a significant impediment towards adoption of more basic data fusion approaches.
Features relevant for WAA of UXO contamination were successfully obtained from the various data sources. Automatic crater extraction from the LiDAR data was successful. An algorithm based on the circular Hough transform was able to extract a value of four meters as the characteristic diameter of craters at the Pueblo site. Further, a pattern recognition algorithm based on the morphology of the craters was developed to locate them in the LiDAR data. This information was then converted into a feature map describing the density aspect of craters. Feature maps describing the intensity and quality aspects to craters provide data fusion algorithms with additional discriminatory information.
Feature extraction algorithms were also developed for magnetometry data. Due to the minimal geomagnetic features at the Pueblo site, a simple threshold proved effective in eliminating geologic background, which is expected to be more significant at other ESTCP Wide Area Assessment Pilot Program (WAAPP) sites. A pattern recognition algorithm was developed to separate ordnance-related signal from the ferromagnetic background of man-made structures. This information was then converted to a feature map describing the density aspect of ordnance-related material. Separate feature maps describing the intensity and morphology aspects of the ordnance-related and man-made components of data are also expected to provide additional discriminatory information for data fusion. Methods for generating feature layers from extracted features for input to data fusion were developed and implemented. For each input feature set, a corresponding feature intensity map and specification of a functional relationship between a feature’s intensity and the likelihoods for or against the presence of UXO that are supported by the feature’s intensity values. Heuristic, Bayesian, and Dempster-Shafer theoretic algorithms for combining evidence presented in feature layers were investigated as possible engines for a UXO WAA data fusion framework prototype. These were implemented as MATLAB code and evaluated with feature layers generated from both the Pueblo and Kirtland site data acquired by performers in the ESTCP WAAPP.
The Dempster-Shafer approach, with its ability to quantify uncertainty about evidence, was shown to be the most appropriate data fusion strategy for the UXO problem and proved to be the most successful of the three. The ability to incorporate heuristic rules regarding specific dependencies between input feature layers into the Dempster-Shafer based data fusion framework prototype was described and demonstrated utilizing two specific examples. The first demonstrated a reduction of false positive indications of UXO by utilizing a feature layer comprised of manually identified man-made structures to selectively block magnetometry-derived features. The second demonstrated an adjustment of the impact of magnetometry-derived features on the output assessment of UXO to accurately reflect the uncertainty associated with increased magnetometry data sparseness in some areas of the helimag survey. The prototype data fusion framework developed was able to delineate areas of likely contamination while providing reasonable estimates of the likelihood of that contamination given supporting observational evidence and a priori knowledge. Preliminary results were compared with limited ground truth data available at the Pueblo site and agreed well.
The key theoretical advantage of a data fusion approach to WAA is the ability to reduce false positives while retaining high detection rates. The framework described is flexible, tolerating missing data and allowing multiple configurations of potential input data streams, as well as scalable, allowing new data streams to easily be included in the assessment. Further, the impact of available and new data streams on the output can be readily quantified. One challenge is that the structured input methodology requires the specification of each feature layer’s relationship to the presence or absence of UXO. However, the input methodology allows specification to be accomplished in a highly flexible manner. The user has the ability to input specifications that vary from simple, intuitive estimations based on expert knowledge to detailed functional relationships based on empirical evidence of sensor performance. Thus, the data fusion framework is capable of utilizing all the information and observation evidence available, without necessarily requiring that the exact same inputs be present for assessment. This flexibility is an important feature of the data fusion approach as it is expected that, for a number of reasons, it will rarely be the case that exactly the same types or quality of data will be available for analysis each time a wide-area UXO assessment is performed.