| This dissertation addresses the problem of detecting targets in clutter background using reflective hyperspectral imagery in real time surveillance and reconnaissance applications. The basic problem is to separate low spatial resolution or sub-pixel targets from clutter background in hyperspectral imagery in real time. The targets are military vehicles, such as tanks, and missile launchers. The clutter includes natural terrain and discrete non-target-sized objects such as roads, and buildings.; The new detection algorithms developed in this dissertation are based on the unique characteristics of the spectral signatures inherent in the hyperspectral data and requires no modeling for target or background characteristics. One algorithm fully utilizes the maximum number of bands available. The other uses the sequential approach utilizing fewer optimum bands. The state-of-the-art advancements of these new detection algorithms include the capabilities to (1) handle high throughput, real-time, sub-pixel, hyperspectral target detection; (2) minimize the effect of variation in sensor calibration, system artifacts, and atmospheric conditions; and (3) yield high probability of detection and low false alarm rate as compared to other classical detection approaches.; This dissertation presents a quantitative detection performance including false alarm rate of these new hyperspectral detection algorithms. Results for different spatial resolutions and different times of day are presented, and compared with the results of other classical detection approaches. Trade-off studies between the probability of detection and false alarm rate versus the number of reflective bands, and the computational complexity based on these new detection algorithms are also presented. In addition, class separation based on the dominant feature of target and clutter distributions versus specific spectral regions is also examined. |
