Identification of unexploded ordnance from clutter using neural neworks

The largest costs associated with subsurface Unexploded Ordnance (UXO) remediation are associated with removing non-UXO. Discrimination between UXO and non-UXO is important for both cost and safety reasons. A neural network was developed to distinguish between UXO and non-UXO clutter using TEM data. There are two stages for the learning process of neural network, training and validation. A synthetic dataset was created using actual acquisition configurations, with varying amounts of random noise. This dataset included 934 UXO targets representing 7 different UXO types, and 789 clutter objects based on four templates with varying size and random asymmetry. The results show 97% accuracy for correctly classifying clutter, and 97% accuracy for correctly classifying UXO. The level of success for classification is based on the classification Receiver Operating Characteristic (ROC) curves. The ROC curve represents the relationship between UXO classified correctly (Hit rate) versus clutter miss classified (False alarm).