Through the ESTCP Wide Area Assessment Pilot Program, three large ranges were evaluated to demonstrate whether, by using modern unexploded ordnance (UXO) survey and data analysis approaches, large portions of the ranges could be removed as areas of concern for UXO contamination. Highly production efficient technologies—helicopter-based magnetometer arrays, Light Detection and Ranging (LiDAR), orthophotography, and synthetic aperture radar—were identified to survey the ranges. In addition, transect surveys and spot area surveys using denser and more sensitive detection from vehicular-towed magnetometer arrays were undertaken to identify target impact areas and areas likely to be ordnance-free. Following these studies, approximately 1,000 targets from each site were remediated to verify the analysis procedures and predictions.
The objective of this project is to apply a new highly efficient and automated data analysis approach to data collected in wide area surveys. Using a small set of ground truth data, researchers will implement an inductive machine learning process using image-based analyses to develop a Target Model, first to select likely magnetic anomalies, then to rank and classify their probability of being UXO.
The machine learning process takes place using learning and post-processing parameter selection from pull-down menus in the commercial software utility, Feature Analyst, operating as an extension to the master utility ArcGIS. Using ensembles of artificial neural networks, the Target Picker continues the training cycle based on known positive and known negative examples, until it converges on a set of regions of interest (ROI) to pass on to the classifier. The classifier begins the process again using the ROI shape functions generated by the Target Picker. The classifier uses the full range of visual attributes (color, texture, pattern, spatial association, shape, and size). However, the primary focus of the classifier is on image shape and size. The shape and size parameters are intentionally suppressed in the target picker cycle. The Target Classifier returns the full list of ROIs ranked by their probabilities of being UXO. To test the approach developed under SERDP projects MR-1322 and MR-1455, this project entails a four-step process, beginning with small training sets involving common vehicular and airborne data, and extending the learning process to the entire site being investigated. In the final step, researchers will incorporate both LiDAR and orthophotography information into a joint analysis with the magnetometry data.
Currently, the most successful target selection and classification approach for UXO survey data relies on human-in-the-loop processing and decision making. This project will streamline the process by creating an automated, efficient approach that provides consistent predictions based on a small set of ground-truthed training data.