A Physics-Guided 3-D Orientation Estimation of Metallic Cylinder Using Multipolarization GPR

Metallic cylinders, such as steel rebars, are commonly encountered targets in ground-penetrating radar (GPR) investigations. The effectiveness of most existing GPR algorithms applicable to such cylindrical targets is inherently reliant on a key geometric premise that the survey line maintains a known, fixed angle relative to the target's longitudinal axis. In real-world applications, if the actual angle between the survey line and the target deviates substantially from the assumed value, the performance of these algorithms can degrade considerably, potentially leading to their failure. Thus, it necessitates a precise estimation of the target's orientation prior to applying these methods in practice. Deep learning, known for its powerful nonlinear fitting capability, has recently been employed to estimate the orientation of these cylindrical targets. However, current deep-learning-based approaches often treat multipolarization GPR B-scans directly as images. Although this strategy can achieve competent performance, it generally suffers from limited interpretability and poor adaptability to variations in the target radius. Different from conventional image-driven approaches, the contribution of this work is to explicitly incorporate electromagnetic-scattering priors into the representation stage and to construct a physics-guided learning framework for 3-D orientation estimation. Starting from the scattering mechanism of a metallic cylinder under multipolarization illumination, we identify which components of the echo remain informative for orientation estimation under radius variation, and then design handcrafted features that jointly preserve phase cues, time-delay information, and hyperbolic geometry while suppressing radius-sensitive amplitude fluctuations. Therefore, the proposed contribution does not lie in a particular network backbone or regression architecture by itself, but in the physically grounded feature representation that bridges multipolarization GPR echoes and orientation regression in an interpretable manner that is robust to radius variation. The effectiveness and remarkable generalization capability of the proposed approach are rigorously validated through both simulated and laboratory experiments.