Small earthquake location via machine learning with insufficient data

Abstract A comprehensive earthquake catalog plays a crucial role in enhancing our understanding of earthquake activity and generation mechanisms. However, due to the low station density and data quality limitations, numerous small earthquakes remain undetected and unlocated. Traditional seismic location methods based on travel time and waveform analysis may be ineffective for such events, while traditional single-station location methods require high signal-to-noise ratio (SNR) data. Overcoming this challenge and improving the detection and location of these small seismic events is crucial. To address the need for locating small seismic events with low SNR, we propose a novel workflow for seismic station networks, leveraging machine learning single-station location method. The method comprises distance and azimuth neural networks. We pre-train the models on a global dataset (STEAD) and fine-tune them using local datasets (INSTANCE, Italy and Sichuan, China). The local datasets are utilized to assess the performance of models, while the Sichuan dataset is also specifically used to evaluate the entire workflow. We evaluate the proposed machine learning methods using the Sichuan, China, testing dataset and achieve a mean absolute error of approximately 3.0 km for epicenter distance and 22.0 degrees for back-azimuth. Extending the application of the models to regions like Yunnan, China, and Italy generates reliable estimates of the spatial distribution of seismic events. Crucially, the workflow incorporates both spatial constraints of the station locations and the constraints from waveforms recorded by the stations, leading to improved location accuracy compared to using less reliable azimuth estimates. The workflow successfully located 6321 seismic events in the Sichuan, China, testing dataset, which is approximately 1.4 times more than the number reported in earthquake catalogs, further complementing seismic activity. Graphical Abstract