Deep learning processing and interpretation of ground penetrating radar data using a numerical equivalent of a real GPR transducer
Ground-Penetrating Radar (GPR) is a popular non-destructive electromagnetic (EM) technique that is used in diverse applications across different fields, most commonly geophysics and civil engineering. One of the most common applications of GPR is concrete scanning, where it is used to detect structural elements and support the assessment of its condition. However, in any GPR application, the data have no resemblance to the characteristics of targets of interest and a means of extracting information from the data regarding the targets is required. Interpreting the GPR data, to infer key properties of the subsurface and to locate the targets is a difficult and challenging task and is highly dependent on the processing of the data and the experience of the user. Traditional processing techniques have some drawbacks, which can lead to misinterpretations of the data in addition to the interpretation being subjective to the user. Machine learning (ML) has proven its ability to solve a variety of problems and map complex relationships and in recent years, is becoming an increasingly attractive option for solving GPR and other EM problems regarding processing and interpretation. Numerical modelling has been extensively used to understand the EM wave propagation and assist in the interpretation of GPR responses. If ML is combined with numerical modelling, efficient solutions to GPR problems can be acquired. This research focuses on developing a numerical equivalent of a commercial GPR transducer and utilising this model to produce realistic synthetic training data sets for deep learning applications. The numerical model is based on the high-frequency 2000 MHz "palm" antenna from Geophysical Survey Systems, Inc. (GSSI). This GPR system is mainly used for concrete scanning, where the targets are located close to the surface. Unknown antenna parameters were found using global optimisation by minimising the mismatch between synthetic and real responses. A very good match was achieved, demonstrating that the model can accurately replicate the behaviour of the real antenna which was further validated using a number of laboratory experiments. Real data were acquired using the GSSI transducer over a sandbox and reinforced concrete slabs and the same scenarios were replicated in the simulations using the antenna model, showing excellent agreement. The developed antenna model was used to generate synthetic data, which are similar to the true data, for two deep learning applications, trained entirely using synthetic data. The first deep learning application suggested in the present thesis is background response and properties prediction. Two coupled neural networks are trained to predict the background response given as input total GPR responses, perform background removal and subsequently use the predicted background response to predict its dielectric properties. The suggested scheme not only performs the background removal processing step, but also enables the velocity calculation of the EM wave propagating in a medium using the predicted permittivity value. The ML algorithm is evaluated using a number of synthetic and measured data demonstrating its efficiency and higher accuracy compared to traditional methods. Predicting a permittivity value per A-scan included in a B-scan results in a permittivity distribution, which is used along with background removal to perform reverse-time migration (RTM). The proposed RTM scheme proved to be superior when compared with the commonly used RTM schemes. The second application was a deep learning-based forward solver, which is used as part of a full-waveform inversion (FWI) framework. A neural network is trained to predict entire B-scans given certain model parameters as input for reinforced concrete slab scenarios. The network makes predictions in real time, reducing by orders of magnitude the computational time of FWI, which is usually coupled with an FDTD forward solver. Therefore, making FWI applicable to commercial computers without the need of high-performance computing (HPC). The results clearly illustrate that ML schemes can be implemented to solve GPR problems and highlight the importance of having a digital representation of a real transducer in the simulations.