Convolutional and Long Short-term Memory Neural Network for Earthquake Detection

Abstract

Earthquake detection is a foundation and critical work link in the study of seismology.Traditional detection algorithms have the disadvantages of low detection sensitivity, low computational efficiency, and poor general applicability.In this paper, we develop a novel earthquake detection method based on convolutional neural network (CNN) and long short-term memory network (LSTM).It extracts the high-level features in the seismic signal and learns the time-frequency characteristics of the main phase by recording the three component data on a single station.We train the network using 4932 waveform windows (2236 positive windows and 2696 negative windows) recorded in Wenchuan, China, with a window size of 30s.Using the trained model to test a continuous waveform for one day, compared to the long-short window energy ratio method (STA/LTA), all manually selected seismic events were successfully detected.To explore the impact of different window sizes and LSTM layers on the detection results, we use a larger dataset (derived from Oklahoma, USA) for network training.The test results show that our method not only has a good generalization ability for cataloging events, but also detects micro-seismic events that are not included in the catalog.The detection accuracy of cataloging events reaches 100%.Our results indicate that this method has fast, efficient and scalable superior performance in earthquake detection.