Surface-wave dispersion curves extraction method from ambient noise based on U-net++ and density clustering algorithm

Abstract

It is crucial to accurately and rapidly extract dispersion curves of surface waves from ambient noise recordings for inverting the subsurface shear (S) wave velocity structures. Conventional manual picking methods are labor-intensive and affected by subjective factors related to seismologists. The existing deep learning methods can automatically and efficiently extract dispersion curves from dispersion spectrograms with a single branch. However, extracting dispersion curves from phase-velocity dispersion spectrograms with multiple branches remains challenging due to complex label-making processes and extensive label-preparation times. Notably, phase velocities typically display a positive correlation with periods and are slightly higher than group velocities. Given the significance of this empirical relationship, we present an intelligent surface-wave dispersion curves extraction method based on U-net++ and density clustering algorithm. Initially, guided by domain knowledge that dispersion curves are smooth, a global searching method is employed to automatically label group-velocity dispersion curves from group-velocity dispersion spectrograms. Subsequently, the group-velocity dispersion curves undergo transformation into probability images of the curves using a Gaussian function. The U-net++ then nonlinearly converts group-velocity dispersion spectrograms into probability images of group-velocity dispersion curves within a high-dimensional space. Following this, the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm is applied to obtain multi-mode phase-velocity dispersion curves from phase-velocity dispersion spectrograms. We then remove invalid dispersion values in the multi-mode phase-velocity dispersion curves based on the empirical relationship between phase and group velocities, ultimately obtaining the phase-velocity dispersion curve corresponding to the group-velocity dispersion curve. Our proposed method has been tested using synthetic data and 3D real-world data collected near Lake Chao in Anhui Province. The test results show that our approach effectively extracts both group-velocity and phase-velocity dispersion curves. Inverted S-wave velocity structures based on extracted dispersion curves in real-world data can characterize the Tan-Lu fault zone, demonstrating the effectiveness of our proposed method.