Integrated Feature Investigation and Classification Methods for Discrimination of Subsurface Objects in GPR Imagery

Ground-penetrating radar (GPR) serves as a valuable sensor for nondestructive imaging of shallow underground objects, with diverse applications ranging from geological surveys to the detection of buried objects, including mines. This research addresses the challenges of classifying underground objects using B-scan GPR imagery, focusing on the discrimination between surrogate mines (both metal and plastic) and nonmine objects such as a metal coke can and a plastic box. Classifying objects based on GPR signatures is challenging due to the presence of clutters and the difficulty of suitable feature selection. To address these challenges, this research work proposes an approach that utilizes a topological active net (TAN) segmentation refined with a hyperbola fitting optimization filter to effectively eliminate clutters in 2-D GPR scans. The research delves into the exploration of various combinations of handcrafted features, employing clustering analysis to identify a combination that exhibits a higher correlation with object characteristics. The selected feature combination is then applied to diverse classifiers like
${K}$
-nearest neighbor (KNN), support vector machines (SVMs), and decision trees to assess their discriminative ability. Experiments are conducted using both synthetic and laboratory-measured GPR images to comprehensively evaluate the efficacy of the proposed solutions.