Predictive Clustering for IED Defeat

Proliferation of sensors, particularly on board unmanned platforms, is quickly enabling persistent surveillance of areas of military interest. The ability to collect and store such surveillance data over large areas and long time periods provides a rich database to mine for patterns of IED activity, but current data mining is human-intensive and limited to forensic analysis of IED events to identify those who emplace and fabricate the IEDs. We propose to extend current capabilities to predictive analysis by developing an automated tool which predicts IED events (locations and times) based on historical data, to estimate patterns, and on real time data, to detect and extrapolate emerging instances of these patterns. Our predictive clustering algorithm computes a list of predicted threat events based on real time data, prioritized by likelihood and characterized by statistical uncertainty of location and time of each event. Complementing the predictive clustering algorithm, a forensic clustering algorithm processes historical sensor data and threat events to estimate parameters in the predictive clustering model. The Phase I effort will develop prototype predictive and forensic clustering algorithms and evaluate them using simulated data to determine performance operating characteristics and using realistic historical data to evaluate military utility.