The use of Ground Penetrating Radar and artificial intelligence for automated railway trackbed stratigraphy and Ballast Fouling assessment

Intrusive trenching and coring remain the reference for railway trackbed diagnosis but lack coverage and repeatability. This paper proposes a hybrid GPR–AI framework that automates the detection of dielectric interfaces and the estimation of ballast permittivity and thickness. Synthetic FDTD simulations are used to evaluate Mask Region-based Convolutional Neural Network (Mask R-CNN) for interface segmentation and XGBoost (gradient-boosted trees)/Support Vector Regression(SVR) for layer-wise regression. Results on controlled data confirm high interface detection accuracy (IoU ≈ 0.81) and robust estimation of shallow dielectric parameters ( R 2 > 0 . 9 ), while sequential conditioning markedly improves deeper-layer predictions. Validation on field measurements acquired with a broadband (40–3000 MHz) GPR antenna array demonstrates good transferability of the methodology, with reliable stratigraphy reconstruction and dielectric-based material attribution along an operational track section. The framework unifies stratigraphy and fouling assessment in a single automated workflow, offering a scalable and interpretable alternative to invasive methods and paving the way for predictive maintenance at the network scale.