Deep learning for the localization of seismic sources based on synthetic full waveform and wavelet de-noising

Abstract

Seismic source localization is an essential technique for the study of earthquakes. Accurate seismic source localization is important in seismic risk assessment. Various machine learning-based methods for earthquake monitoring and source localization have been proposed, along with the development of source localization techniques. However, these methods require a large amount of historical data for training, and acquiring the required data using monitoring stations may take years or even decades. Moreover, the acquired data often contain various seismic noise types that can affect the calculation results. To address this problem, we combine wavelet de-noising with convolutional neural network (CNN) to achieve fast source localization without any historically cataloged events. The results show that adding the wavelet de-noising technique improves the proposed model. In addition, provided that the regional model is known a priori, the method has a wide range of applications. For example, it can be applied to scenarios such as rock bursts in mines, microseismic events generated by mining, or big earthquakes. Based on this approach, we also have the potential to build a picking-free, non-historical catalog, noise-robust, and fully automated location method.