Overview
Ground-penetrating radar (GPR) data is noisy, and conventional inversion is expensive and slow. This project trains neural networks to approximate the inversion directly from the radar trace.
Pipeline
- Radar trace preprocessing and normalization
- Network training with a physics-aware loss to keep reconstructions physical
- Comparison of predicted inversion against ground-truth reference data
Outcome
Prediction examples are available in the repository, and the approach illustrates how deep learning can speed up an otherwise expensive inverse problem.