Automatic picking of multi-mode surface-wave dispersion curves based on machine learning clustering methods

The surface-wave analysis method is widely adopted to build near-surface shear-wave velocity structure. Automatic picking of surface-wave dispersion curves is a problem that needs to be solved, especially when dealing with a large amount of data. In this paper, we proposed an automatic picking method of multi-mode surface-wave dispersion curves based on unsupervised machine learning methods. Firstly, seismic data is transformed to two-dimensional dispersion image by dispersion imaging method. The discrete dispersion image points are separated into two clusters: dispersion energy and background noise by Gaussian mixture model clustering method. Then the dispersion energy points are projected into frequency-velocity domain, and are distinguished as different surface-wave modes by the density-based spatial clustering of applications with noise algorithm. Finally, multi-mode dispersion curves are obtained by searching for the local peaks in different mode areas of surface waves. Particle filter is used to smooth the picked dispersion curves to eliminate the influence of noise. The synthetic tests show that the automatically picked dispersion curves can match the theoretical dispersion curves well. We also applied the proposed method to the picking of dispersion curves of two field data. The automatically picked dispersion curves are applied to the subsequent inversion. In the second field data, the one-dimensional inversion results are interpolated to obtain the near-surface three-dimensional velocity model of the work area. The inversion results are in good agreement with borehole data, which further proves the accuracy and efficiency of the automatic picking method for multi-mode dispersion curves.

Automated summary

The paper proposes an automatic picking method for multi-mode surface-wave dispersion curves based on unsupervised machine learning, which transforms seismic data, clusters dispersion points, and searches for local peaks in mode areas. The results show that the automatically picked dispersion curves match theoretical curves, and are successfully applied in inversion and 3D velocity model construction.