Reinforcement Learning-Driven Sparse Frequency Analysis for Accurate Underground Pipe Detection in GPR Scans

Abstract

Ground penetrating radar (GPR) is an essential tool in the modern subsurface investigations due to the effectiveness and non-destructive potential of detecting objects. GPR is commonly used to identify and locate any subsurface structure and anomalous body. Ground surface variability, signal clutter and noise often hinder the use of GPR in the identification of underground pipes. In that regard, this paper presents a Sparse Frequency Transform-based Auxiliary Reinforcement Model that is capable of providing accurate and effective pipe detection inside GPR data. Image restoration to increase the quality of GPR scans is the starting point of the methodology where a sparse frequency transform is then used to isolate salient features and eliminate extraneous background signals. Such features are then fed into an auxiliary reinforcement model, and gradually, it enhances the detection accuracy by means of a rewarded-based learning process. Experimental evidence proves that the suggested model is better in relation to the traditional methods, including RefineNet, YOLOv4 and YOLOv8, in terms of accuracy, sensitivity, specificity, and precision. Combining sparse representation and reinforcement learning provides a solid platform on useful subsurface utility mapping and infrastructure evaluation.