Abstract
Surface wave is an energy-rich component of the seismic wavefield and has been widely employed in understanding underground structures due to its dispersive nature. One key work in improving the accuracy of dispersion curve measurement is selecting proper cycles and valid frequency ranges. Although manual selection could provide high-quality results, it is hardly possible to handle the explosive growth of seismic data. Conventional automatic approaches with the ability to handle massive datasets by their statistical features require prior assumptions and choices of parameters. However, these operations could not keep away from biases in empirical parameters and thus could not assure high-quality outputs, which might deteriorate the resolution of seismic inversion. To make good use of the waveform information, we develop a deep-learning-based neural network called 'Surf-Net'. It extracts and selects the surface-wave dispersion curves directly from the waveform cross-correlations (CC) and distance information rather than from frequency-time transformed images or pre-extracted dispersion curves. Taking the velocity measurement task as an arrival time picking problem, Surf-Net is designed to output multiple-channel probability distributions in the time domain for target frequencies, which peak at the arrival times of valid frequencies and remain close to zero elsewhere. We train and test Surf-Net using observational data manually obtained from seismograms recorded by a regional network in Northeast China and synthetic data based on a global seismic velocity model. By comparing Surf-Net with the conventional method in both dispersion curves and inversion results, we show Surf-Net's remarkable performance, robustness and potential for providing high-quality dispersion curves from massive datasets, especially in low frequencies.