Managing within-class target variability in SAR imagery with a target decomposition model

Modifications in a target's signature caused by minor variations in the target can severely limit the robustness of most techniques designed to find targets in synthetic aperture radar (SAR) imagery This lack of robustness is a pressing issue for SAR. Part of the problem, perhaps, is that targets have been considered as single units and hence features have been extracted with the assumption of radar cross section (RCS) stability everywhere in their footprints. The author demonstrates the advantage of not considering targets as single units; instead, a target is considered in terms of a spatially decomposed model. The model exploits multiple RCS regions in targets. The key features in the model are that (1) targets can be spatially decomposed into useful RCS regions, (2) pixel values in the target are independent random variables in quarter power, (3) the metric /spl sigma//sub i///spl mu//sub i/ determines the level of RCS stability in these regions, and (4) the random variables pertaining to stable regions are Gaussian distributed. The utility of this model is illustrated through a study case involving a serious problem of signature variability.