DisperNet: An Effective Method of Extracting and Classifying the Dispersion Curves in the Frequency–Bessel Dispersion Spectrum

Abstract

ABSTRACT The subsurface shear-wave structure primarily determines the characteristics of the surface-wave dispersion curve theoretically and observationally. Therefore, surface-wave dispersion curve inversion is extensively applied in imaging subsurface shear-wave velocity structures. The frequency–Bessel transform method can effectively extract dispersion spectra of high quality from both ambient seismic noise data and earthquake events data. However, manual picking and semiautomatic methods for dispersion curves lack a unified criterion, which impacts the results of inversion and imaging. In addition, conventional methods are insufficiently efficient; more precisely, a large amount of time is required for curve extraction from vast dispersion spectra, especially in practical applications. Thus, we propose DisperNet, a neural network system, to extract and discriminate the different modes of the dispersion curve. DisperNet consists of two parts: a supervised network for dispersion curve extraction and an unsupervised method for dispersion curve classification. Dispersion spectra from ambient noise and earthquake events are applied in training and validation. A field data test and transfer learning test show that DisperNet can stably and efficiently extract dispersion curves. The results indicate that DisperNet can significantly improve multimode surface-wave imaging.