Principles of Modern Radar - Volume 2 1891121537

14.4 Synthetic Aperture Radar 64714.4.4.2 Template LibrariesIn a template-based approach to ATR, templates are created from collected SAR imageryand compared to the test data. The template can be a spatial signature for one specific targetgeometry or state, or it can be a linear combination of registered signatures over severalconditions [42]. One example of combining signatures over several geometries is the MITLincoln Laboratories ATR approach in which target image templates were generated byaveraging five images acquired over 5 ◦ of aspect angle [7].14.4.4.3 Adaptive FeedbackRather than using training data based on supervised learning or solely on a priori imagecollections, a method for using feature-based classification in conjunction with truthfeedback to improve ATR learning and classification over the course of a mission can beemployed [43]. By doing so, operational conditions that are not included in the trainingdata can be addressed. In the Air Force Research Laboratory’s (AFRL) AdaptSAPS systemand Sensor and Data Management System (SDMS), truth feedback is assumed to comefrom human analysts with access to other information sources [43]. The quadratic distancemeasure determines which class an object belongs to and changes adaptively by refiningthe feature vector mean and covariance as more data is received during the mission [43].14.4.4.4 ModelsModel-based ATR generates features from mathematical models of the sensing processand target geometry, which can then be matched to features from unknown signatures[42]. One well-documented model-based algorithm is the PEMS modules of the MSTARprogram, which follows the hypothesis-and-test approach for hypothesizing a target type,extracting the appropriate feature set from models, performing matching, and determiningthe likelihood score [42]. This iterative ATR subsystem uses high-fidelity models of targetsto refine and verify target classification [4], as described earlier.The European Aeronautic Defence and Space (EADS) Company has created a radarimage generator called SAR-VIEW which consists of a terrain database, feature mapconstruction, height map construction, radiometric expectation values, CAD models fortargets, and formation of a synthetic SAR image. First, a set of terrain data is extracted fromthe terrain database. Using pre-defined requirements for the resolution of the image andknowledge about the platform heading, a landscape module transforms and interpolatesthe terrain for the desired conditions to produce a map of height values and a map of landcover types. The expected clutter RCS at each pixel of the map can then be calculatedfrom ground and slant ranges. A combination of geometric and physical optics is used tofind radar cross sections from 3D CAD models of man-made targets. After the targets areintegrated with the landscape, pixel correlation and speckle effects are added to give thefinal synthetic SAR image. Additional approaches for improving the image realism canbe employed in this system. These images are useful for ATR training and for integrationinto combat flight simulators [44].14.4.4.5 Persistent ScatterersPersistent scatterers have been shown to exist over nearly the entire range of possible targetorientations in real and synthesized (XPatch) SAR images [45]. They can be extracted fromhigher resolution SAR images or from Xpatch electromagnetic signature predictions of3-D CAD models [46].