Contextual Features and Bayesian Belief Networks for Improved Synthetic Aperture Radar Combat Identification

Automatic target recognition (ATR) considers detecting, classifying and identifying different manmade objects for combat identification (combat ID). Synthetic aperture radar (SAR) imagery offers a high-resolution, and weather invariant sensing modality for combat ID. However, SAR and the associated high-range resolution (HRR) profiles of even the same target will have different signatures when viewed from different angles. To overcome this challenge, exploitation of contextual information is considered herein. Contextual information is considered consistent with content-based image retrieval (CBIR) methods and considers features describing the shape of a detected target. Data from a wide range of aspect and depression angles is used to train pattern recognition algorithms, and data from unexplored and larger depression angles is used for testing. This research fuses segmentation algorithms and multivariate analysis methods to extract contextual features from SAR imagery. The contextual features are used in conjunction with HRR features improve classification accuracy at similar or extended operating conditions. Three classifiers are considered: baseline template matching, a probabilistic neural network (PNN), and Bayesian belief networks (BBN). The proposed method is shown to have statistically significant classification accuracy improvements over baseline methods in literature.