Object detection using transformed signatures in multitemporal hyperspectral imagery

Changes in atmosphere, ground conditions, and sensor response between multitemporal airborne imaging sessions have limited the use of fixed target hyperspectral libraries in helping to identify targets in heterogeneous (cluttered) backgrounds. This hyperspectral target signature instability has resulted in using anomaly detection algorithms to detect targets in real time applications. The anomaly detection algorithms, however, have not detected targets at sufficiently low false alarm rates. This study examines mathematical transforms of target spectral signatures. Specifically this study uses statistical information regarding background clutter taken from one long-wave infrared (LWIR) hyperspectral (8-12 μm) airborne imagery flown on one day, to find the target spectral signature flown on another day (with significantly dissimilar weather conditions). The transforms use overlapping regions in the two data sets but without subpixel level registration. This work analyzes image cubes collected during the November 1998 Hyperspectral Day/Night Radiometry Assessment (HYDRA) data collect. The transformed signatures used in matched filter searches successfully find targets (even targets nearly covered) with low false alarm rates (<1 FA/kilometer²) and remained sensitive to targets using a reduced number of pixels in the overlap region. This work has demonstrated the transformation of target spectral signatures to search for candidate targets using multitemporal hyperspectral images without requiring accurate geo-registration