Micro-seismic events detection and its tectonic implications in Northeastern Hainan Province

Abstract

In this paper, we combine the U-net-based phase picking method (PhaseNet) with Graphics Processing Unit-Based Match and Locate technology (GPU-M&L) and a deep-learning-based seismic signal de-noising method (DeepDenoiser) as a workflow for automatically extracting micro-seismic information from continuous raw seismic data. PhaseNet is first used to detect missed seismic phases by scanning through the 5-year continuous waveform data recorded at five broad-band stations in Hainan province. Then Rapid Earthquake Association and Location method (REAL), VELEST program (1-D inversion of velocities and hypocenter locating) and HypoDD (a double-difference locating method) are applied to associate seismic phases with events and to locate, respectively. This initially established catalogue can be served as the template for the following match-filter work. We choose events with a high signal-to-noise ratio (SNR) as templates and apply GPU-M&L to detect more small earthquakes which are difficult to pick by routine methods due to the low SNR. Then, a deep learning-based noise reduction technique named DeepDenoiser is applied to extract seismic signal from noise to provide a better picking of arrival time and then to improve the relocation effects. Finally, we use HypoDD to relocate these events with P- and S- wave arrival times picked by PhaseNet. Compared with the five events listed in the China Earthquake Networks Center routine catalogue, in this study, we detect and locate 977 earthquakes by following the above procedure. Our relocation results illustrate quite a complex distribution pattern of events due to the complicated fault system in the northeastern part of Hainan Province.