Automatic Target Recognition for Hyperspectral Imagery

Automatic target detection and recognition in hyperspectral imagery offer passive means to detect and identify anomalies based on their material composition. In many combat identification approaches through pattern recognition, a minimum level of confidence is expected with costs associated with labeling anomalies as targets, non-targets or out-of-library. This research approaches the problem by developing a baseline, autonomous four step automatic target recognition (ATR) process: 1) anomaly detection, 2) spectral matching, 3) out-of-library decision, and 4) non-declaration decision. Atmospheric compensation techniques are employed in the initial steps to compare truth library signatures and sensor processed signatures. ATR performance is assessed and additionally contrasted to two modified ATRs to study the effects of including steps three and four. Also explored is the impact on the ATR with two different anomaly detection methods