We implemented a Deep Learning algorithm to estimate the subsurface EM velocity field from common offset GPR profiles. The Deep Learning approach is based on a Bi-Directional Long Short-Term Memory (LSTM) Neural Network (NN) architecture trained on simple synthetic profiles randomly generated.
The trained network is then applied to each A-Scan of 2D or even 3D GPR datasets. We trained the network on a synthetic dataset with different numbers of reflectors, wavelets, Signal-to-Noise ratios. The application of the network to synthetic and field data successfully predicts the velocity model and provides a computationally effective alternative to classic methods.
