A Deep Learning framework for Ground Penetrating Radar

Machine learning (ML) is becoming a more frequently used approach to deal with GPR and other electromagnetic problems, which due to the complexity of the data, require new more complex solutions. We have developed an ML framework to provide solutions to specific GPR applications and scenarios. The ML tools utilize neural networks (NNs) and a large training set originating from simulations that include a digital twin of a real GPR transducer. The applications investigated are background removal, automatic estimation of the background bulk permittivity in conjunction with a reverse time migration (RTM) scheme that utilizes the ML outputs and is applied to reinforced concrete slab scenarios. The schemes are validated using both synthetic and real data, showing a very good accuracy and demonstrating the success of the ML algorithms. Although, this ML framework is applicable to certain applications and scenarios, it can be easily extended to other classes of problems.