A hybrid multi-task 1D CNN and XGBoost approach for ballast layer characterization using GPR data

• Multi-task 1D CNN predicts ballast depth, permittivity, and fouling class. • Model trained on GPR traces with excavation-based ground truth from 4 sites. • CNN achieved high accuracy across all tasks with real-time inference speeds. • Hybrid XGBoost model improved permittivity estimation using spectral features. • Feature importance analysis confirms synergy of learned and engineered inputs. Ground-penetrating radar (GPR) enables rapid, non-invasive subsurface inspection; however, current practices depend on manual interpretation and heuristic thresholding. This paper introduces an automated framework for ballast layer characterization using a lightweight multi-task one-dimensional convolutional neural network (1D CNN) that jointly estimates ballast thickness, relative dielectric permittivity (ε r ), and fouling level from raw GPR A-scan signals. Trained on 2,358 labeled traces collected from four active rail corridors with 14 excavation pits, the model achieved a mean absolute error (MAE) of 1.95 cm for thickness, 0.062 for permittivity, and 95.4% classification accuracy. A hybrid model combining CNN-derived features with engineered spectral descriptors via XGBoost further improved thickness (MAE = 1.28 cm) and permittivity (MAE = 0.029), while maintaining a comparable fouling classification accuracy of approximately 95.5%. Feature-importance analysis highlighted the dominance of CNN-derived features, with engineered spectral features such as fast Fourier transform (FFT)-based spectral area and spectral centroid providing complementary information. By automating subsurface condition monitoring, the proposed framework advances intelligent inspection and maintenance management of railway infrastructure, contributing to the broader field of automation in construction and lifecycle asset management