Day/Night Polarimetric Anomaly Detection Using SPICE Imagery

We introduce a novel longwave polarimetric-based approach to man-made object detection that departs from a more traditional direct use of Stokes parameters. The approach exploits the spatial statistics on two coregistered vertical and horizontal polarization components of the images, where differences of spatial second-order statistics in the bivariate space reveal that man-made objects are separable from natural objects while holding invariant to diurnal cycle variation and geometry of illumination. We exploit the invariant feature using the Bayes decision rule based only on probabilities. Experimental results on a challenging data set, covering a 24-h diurnal cycle, show the effectiveness of the new approach on detecting anomalies; three military tank surrogates posed at different aspect angles are detectable in a natural clutter background. These results yield a negligible false alarm rate as the heating components of the tank surrogates were turned off during data collection.