WAVEFIELD DECOMPOSITION FOR DIFFRACTION SEPARATION WITH A CONVOLUTIONAL AUTOENCODER FOR SEISMIC AND GPR DATA

In the recent years, machine learning and artificial intelligence have gained increasing importance across all fields of Earth sciences. In particular convolutional neural networks (CNNs) have proven to be a powerful tool for different types of data analysis, such as pattern recognition, image segmentation, denoising and data reconstruction. In the context of seismic and electromagnetic imaging, the decomposition of the measured wavefield into its reflective and diffractive components is a notorious challenge that has been approached by numerous deterministic approaches. In this study, we generated a large set of synthetic seismic data with randomized features and parameters that consist of three wavefield components - reflections, diffractions and noise. Since our modeling approach allows to generate each component separately, we were automatically provided with labels for the training of a neural network. We trained a convolutional autoencoder with the generated synthetic data and the corresponding labels and subsequently applied the trained neural network to seismic and GPR field data. The promising results show that the neural network appears to be able to correctly identify and map the sought-after wavefield components and thus demonstrates the transfer-learning capabilities of convolutional neural networks.