Deep Learning for Extracting Dispersion Curves

High-frequency surface-wave methods have been widely used for surveying near-surface shear-wave velocities. A key step in high-frequency surface-wave methods is to acquire dispersion curves in the frequency–velocity domain. The traditional way to acquire the dispersion curves is to identify the dispersion energy and manually pick phase velocities by following energy peaks at different frequencies. A large number of dispersion curves need to be extracted for inversion, especially for surveys with long two-dimensional sections or large three-dimensional (3D) coverages. Human–machine interaction-based dispersion curves extraction, however, is still common, which is time-consuming. We developed a deep learning model, termed Dispersion Curves Network (DCNet), that can rapidly extract dispersion curves from dispersion images by treating dispersion curves extraction as an instance segmentation task. The accuracy of the dispersion curves extracted by our DCNet model is demonstrated by theoretical data. We used a 3D field application of ambient seismic noise to demonstrate the effectiveness and robustness of our method. The real-world results showed that the accuracy of the dispersion curves extracted from the field data using our method can achieve human-level performance and our method can meet the requirement of geoengineering surveys in rapidly extracting massive dispersion curves of surface waves.