Robust Machine Learning for UXO Detection

Many of the areas designated for Base Realignment and Closure (BRAC) and Formerly Used Defense Sites (FUDS) are contaminated with Unexploded Ordnance (UXO) that needs to be cleared of before being returned to civilian use. Despite the significant research and development efforts, there is still plenty of room for improving the UXO detection technologies. For EMI and GPR sensors, the predominantly metallic UXO poses less of a problem for detection. The limiting factor is the high false alarms caused by metallic clutter often prevalent at those survey sites that add considerably to the clean-up cost due to the excavation. To mitigate the false alarms, many signal processing methods have been proposed that adapt the sensor signal to local environment, and/or impose UXO specific features to better differentiate UXO from clutter objects. Although many of the methods show good potential in a controlled environment, their actual performances in more realistic field tests often degrade, sometimes significantly, from those achieved in the laboratory settings due to the variability and diversity of UXO and the uncertainty in the UXO locations. In this effort, we propose several approaches for object feature extraction, feature selection, and classification of the detected objects. The main theme of our proposed approaches is to ensure robustness of the resultant UXO classification results under conditions that are not exactly the same as those assumed by the simplified target models or those of the training data.