Seismic event classification based on a two-step convolutional neural network

Abstract

The identification of unnatural earthquake events is one of the tasks of earthquake rapid report. The identification accuracy is of great significance for improving the quality of earthquake catalog and seismological research. In this study, a 7-layer convolutional neural network model was constructed to identify unnatural earthquakes. First, the three-component seismic waveform was input to obtain the waveform image classifier, and then, the time–frequency spectrum of blasting and collapse was input to obtain the time–frequency spectrum classifier. The two classifiers were used to identify natural earthquake, blasting, and collapse. The model was trained and tested using 3386 seismic events of Shandong seismic network from 2017 to 2022. The events identified as blasting by the waveform image classifier were reidentified by the time–frequency spectrum classifier. Finally, the identification accuracy of natural earthquake, blasting, and collapse is 97.50%, 95.87%, and 86.84%, respectively, with an average accuracy rate of 96.13%. The experimental results show that the two-step convolutional neural network can extract the characteristics of seismic signals from multiple angles, which get a good result in seismic event classification.