Double-Difference Tomography with a Deep Learning–Based Phase Arrival Weighting Scheme and Its Application to the Anninghe–Xiaojiang Fault Zone

Abstract

High-precision seismic phase arrivals are a prerequisite for building reliable velocity models with travel-time tomography. There has recently been a growing use of seismic phase arrival data obtained through deep learning techniques in travel-time tomography research. Nevertheless, a significant challenge that has emerged pertains to the assessment of the quality of these automatic arrivals. In this article, we used PhaseNet, a deep learning method, to automatically detect the arrival times of the P wave and S wave of 3086 seismic events recorded by dense seismic arrays, obtaining 87,553 high-quality arrivals. To evaluate the quality of the arrival times subsequently used for travel-time tomography inversion, we applied a weighting scheme that includes both detection probability value and signal-to-noise ratio. This new weighting scheme can effectively reduce the overall travel-time residual by 7%. The weighted data were then used in the double-difference tomography method to invert for the crustal velocity structure of the Anninghe–Xiaojiang fault zone. The resulting new model exhibits a lateral resolution of up to 0.25° and reveals velocity anomalies that exhibit a strong correlation with major geological features and block boundaries. Notably, the presence of low-VP and low-VS in the middle crust of the Ludian–Qiaojia seismic zone suggests the existence of hot and weak felsic rocks, as well as possible fluid presence beneath the seismogenic layer of this area. This study not only validates the practicality of using deep learning-based phase picking arrivals in travel-time tomography but also proposes a new weighting scheme to refine the tomographic velocity models.