Building an initial model for near-surface surface waves based full-waveform inversion with deep learning

Full-waveform inversion (FWI) is capable of providing high-resolution models of the elastic parameters of the subsurface at different scales, including shallow seismic applications. In spite of its vast potential, FWI still has certain challenges that need to be addressed, especially for near surface applications. One of the biggest being defining an initial model. Recent advances in deep learning (DL) algorithms combined with more accessible computing resources allowed us to perform direct shear-wave velocity (Vs) model extraction from surface-wave dominated data. For this approach we developed a convolutional neural network (CNN) with approximately 100,000 data-model pairs generated from random models within a predefined range of velocity variations. Our synthetic tests show that our CNN produced 2-D models comparable with the target model down to 20 m below the ground surface (BGS). We also tested our neural network on field data collected over a known void at about 10 m BGS. While the void could not be identified in the CNN output itself, the output model served as a good starting point to improve data fit for FWI and ultimately model resolution.