Real-time earthquake location via convolutional neural network and data fusion

Abstract:

Locating earthquakes has always been a basic but challenging task in geoscience. This study presents the development of the deep learning-based location framework to realize real-time earthquake location. By using deep learning techniques, we aim to learn the location inversion mapping function directly from the input data and the labels. Our deep-learning framework differs from prior studies by fusing waveform data and the corresponding receiver layout information to locate the epicenter with the uncertainty estimation. The model consists of 3 modules. The first module is a convolutional neural network (CNN) that extracts the features of each earthquake waveform. Module 2 concatenates the produced features with the station layout data, and the concatenated data is fused to produce the vectorized distance from the reference station to the source in module 3. The source location is then determined by the distance vector and the reference station.

To evaluate the capacity of our model, we first use synthetic data generated from a layered velocity model with point seismic sources and receivers. The performance of the model is evaluated by the statistical analysis of the prediction errors. We verify the role of the fusion processor (module 2) by analyzing the model parameters, and this analysis indicates that our model is in line with physical knowledge. By assigning different receiver assembles and reference receivers, the model can locate the earthquake position with the uncertainty estimation. Finally, we applied the model to the microearthquake data from the Newberry geothermal project. The results indicate that the accuracy of the real-time microearthquake locations produced by our model meets the requirement of the practical application requirements.