Training a Seismogram Discriminator Based on ResNet

Selecting appropriate data of good quality is the primary precondition for conducting meaningful analysis in seismological research. The data set used in a study is usually selected by setting a threshold for searching the earthquake catalog to ensure data quality, such as parameters of magnitude or hypocentral distance. While the threshold approach is useful, it cannot guarantee the consistently good quality of the selected seismograms. For that reason, a manual checking process is often required for quality control, which is inefficient and fallible for large data sets. In this study, we develop an automated seismogram discriminator that is capable of selecting quality seismograms from massive events. The discriminator is created based on a machine learning technique using a residual neural network (ResNet), a well-designed architecture in computer image recognition. Three-component seismic records of an earthquake from the catalog are considered as images by the ResNet. Using the images of seismic records, the ResNet can be trained to distinguish between good and poor seismic records. This discriminatory ability is evaluated using a blind testing data set of approximately 20 000 three-component seismic records related to the events that occurred in Sichuan between January 2014 and May 2018. The results show that our seismogram discriminator achieves an accuracy of greater than 95%.