Automatic recognition of targets on very high resolution SAR images

Recently, different very high-resolution synthetic aperture radar (SAR) missions have been launched, but the great potential of SAR systems for intelligence and defense purposes has been only partially exploited until now, because SAR images are much more difficult to interpret by human operators w.r.t. optical ones. In particular, the aspect of targets in SAR images depends dramatically on the relative orientation between line of sight and target. We have devised a methodology and we have developed a prototype for automatic target recognition on SAR imagery. Differently from the few previous studies available in the literature, our approach is based on machine learning techniques applied to the images themselves, possibly after some linear or nonlinear filterings to improve robustness. A SAR simulator and CAD models to create a database of SAR target signatures for the classifier training. The developed prototype was validated on an extended set of images of military vehicles taken by different SAR satellite and airborne systems, with good results confirming the validity of the proposed approach.