Detection of Small Earthquakes by Waveform Envelope Using Machine Learning

ABSTRACT Analyzing seismic data has helped reveal most of our knowledge about the Earth’s interior. However, the number of moderate-to-large earthquakes is limited; and most small earthquakes, although essential for monitoring the dynamic process between major earthquakes, are difficult to detect from earthquake recordings. In this study, we present a detection method that catches the envelope pattern of seismic data in machine learning. The waveform envelope presents reliable features and a low-frequency pattern for any event rather than being dominated by details of noise and signals. We first apply a fully connected neural network to extract data envelopes from recordings and then a convolutional neural network to detect events using the envelopes as input. Our method is tested against the previously published method on seismic data in Japan. The new approach identifies the largest number of small events. In the testing dataset, the precision and recall of the approach for events are 98.73% and 96.54%, respectively, and those for noise are 97.41% and 99.05%, respectively. We demonstrate that the approach performs well in different signal-to-noise ratios (SNRs) and filter frequency band tests. For events with low SNRs (0–6 dB), the detection accuracy of the approach is approximately 10% higher than a published deep learning method. Besides most of the events in the earthquake catalog, the process also detects more events from continuous data that are not documented in the catalog.