Convolutional Neural Network, Res‐Unet++, ‐Based Dispersion Curve Picking From Noise Cross‐Correlations. Issue 11 (11th November 2021)

• Record Type: Journal Article
• Authors: Song, Weibin; Feng, Xuping; Wu, Gaoxiong; Zhang, Gongheng; Liu, Ying; Chen, Xiaofei
• Is Part Of: Journal of geophysical research. Volume 126:Issue 11(2021)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2021-11-11
• Subjects: deep learning -- surface wave -- dispersion curves -- ambient seismic noise -- transfer learning -- neural network; Geomagnetism -- Periodicals; Geochemistry -- Periodicals; Geophysics -- Periodicals; Earth sciences -- Periodicals; 551.1
• DOI: 10.1029/2021JB022027
• Languages: English
• ISSNs: 2169-9313

Abstract

Abstract: Ambient seismic noise cross‐correlation has been widely applied in surface wave tomography at regional to global scales, including for seismic exploration of near‐surface structures. Reliable seismic imaging requires the accurate selection of dispersion curves. However, manual picking has become cumbersome work with the increase in available correlation traces; it is even more difficult when the number of dispersion curves increases by using frequency‐Bessel (F‐J) transform. Here, we show that the neural network Res‐Unet++ can automatically and accurately extract both fundamental dispersion curves and overtones from the F‐J dispersion spectra after training the network. Results show that selected dispersion curves had high accuracies in the synthetic data (greater than 95%). The network could effectively extract both the fundamental and higher modes in real data, and transfer learning improved the adaptability of neural networks for different geological areas. The obtained dispersion curves from the real data agreed well with those acquired manually and were advantageous for generating more effective dispersion points. Plain Language Summary: Seismic surface wave has been widely used in investigating structure of our planet. When it propagates in the underground structure, the velocity of seismic surface wave changes with frequency. To accurately investigate structure of our planet, we need to get the relation between velocity and frequency, and this relation is expressed as dispersion curves, which is recorded in an image. We usually obtain this relation manually, but it is labor‐intensive and time‐consuming. To save time and energy, we use computer to automatically obtain the relation between velocity and frequency. After the processing based on computers' learning, we can obtain the velocities at different frequencies in a short time. The obtained results for synthetic data and real data proved to be reliable, and the accuracies are relatively high, which can help us investigate the structure of our planet. The automatic method can apply to more data from different areas by keeping the computer learning some new data. Key Points: The fully convolutional network Res‐Unet++ is applicable to effectively extract multimode dispersions from dispersion spectra The automatically and manually picked dispersions have significant similarities Using transfer learning has alleviated effectively the inadaptation problem in spectra from different areas