Machine learning based Surface wave mode identification

Dispersion of seismic surface waves, generated by earthquakes or reconstructed from ambient noise correlations, is widely used to study the structure of the Earth, with tomography, or its temporal evolution, with seismic interferometry monitoring. Nearly all of these studies focus on the fundamental mode of either Rayleigh or Love waves. However, as the medium of propagation becomes complex, for example with sedimentary basins, higher modes emerge, and failure to account for them can bias the analysis. Moreover, using multiple modes information from different types of waves helps to get deeper and shallower depth sensitivity and reduces the non-uniqueness of the solutions. Under-utilization of surface wave overtones comes in part from the relative difficulty of identifying them and extracting their dispersion characteristics. Here we propose a new approach, based on machine learning, to identify and pick multiple-mode dispersion curves of surface waves. We compute thousands of synthetic seismograms from different 1D models with clearly labeled and separated modes and generate their corresponding frequency-time analysis. We then combine the unimodal diagrams and seismograms into full multi-modal diagrams and seismograms (plus noise) and train our algorithm on this dataset. The particle motion of the Rayleigh-wave can give crucial help to differentiate the modes since the fundamental mode is expected to be retrograde and the first higher mostly prograde. For this reason, we analyze the particle motion of the correlations to better dissociate the modes and add this a priori information. We apply our procedure on real seismograms obtained from ambient noise cross-correlations from Central California, where the deep sedimentary basin of Central Valley generates strong higher modes.