SDCnet: An Unet with residual blocks for extracting dispersion curves from seismic data

摘要

The ambient noise cross-correlation technique has been widely used to obtain the S-wave velocity structure. Accurate and efficient extraction of dispersion curve from the spectrum energy diagram is extremely important. In this study, a data-driven deep learning network, SDCnet, is proposed for dispersion curve extraction. SDCnet is an improved Unet with a residual block structure. It automatically and intelligently identifies the peak values at different frequencies as a case segmentation task. A trainable upper and lower sampling strategy was introduced into the residual block to improve the ability of feature extraction and avoid network degeneration. Training set from real and synthetic data widely improves the SDCnet's generalization ability. In our study, SDCnet has demonstrated its high accuracy and efficiency in applications of real data from USArray in the western USA. Compared with the traditional model-driven methods, SDCnet can accurately extract dispersion curves without prior information. The network has the advantage of saving labor and time for extracting massive dispersion curves.