Deep Learning-Based GPR interpretation of soil thickness in headwater hillslopes

Abstract

Soil thickness strongly influences eco-hydrological and geomorphic processes, yet conventional measurements such as auger drilling are invasive, labor-intensive, and unsuitable for large-scale surveys. Ground-penetrating radar (GPR) provides a non-invasive alternative, but its manual interpretation remains slow and prone to observer bias. To address this challenge, we developed a fully automated framework that couples a hybrid CNN-Transformer deep learning architecture with optimized signal filtering to predict soil thickness directly from GPR profiles. The convolutional layers extract local waveform features, while the attention mechanism captures long-range dependencies. Using field data from a steep headwater hillslope (H1) in the Taihu Basin, China, we compared five filtering strategies—median, Savitzky-Golay, Gaussian, moving average, and none—and found that median filtering yielded the most accurate results (R² up to 0.92, CCC of 0.96, RMSE near 10 cm). We further identified optimal filter window sizes (61–101 samples) and a training duration threshold (≥500 epochs) that ensured stable and accurate predictions. Cross-site validation on an independent hillslope (H2) without retraining showed that the pretrained CNN-Transformer model achieved the highest R² (0.80), CCC (0.89), and lowest RMSE (11.3 cm), outperforming traditional machine learning models (CNN, MLP, RF, SVM) in transferability. These findings demonstrate that integrating CNN-Transformer architectures with appropriate signal filtering enables scalable, accurate, and objective soil thickness mapping in complex terrain. The proposed approach also holds promise for broader GPR-based subsurface applications, including soil horizon delineation and root system detection.