統計学輪講 第27回

We present a semi-automated method for identifying and classifying different kinds of seismic events recorded in continuous seismograms by a single station. The method first utilizes the well-developed PhaseNet picker to identify events of interest, and subsequently applies a modified Deep Embedded Clustering (DEC) architecture to classify them. DEC is a self-supervised deep neural network capable of learning the salient features of a dataset while simultaneously clustering the dataset using those features, eliminating the need for manual feature engineering and labeled dataset preparation. We apply this workflow to a dataset recorded by station TA.O53A, which is located in Harrison County, Eastern Ohio. The dataset contains several well-studied hydraulic fracturing induced earthquake sequences and numerous blasting events. Using the proposed method, we can separate earthquakes from blasting events in the dataset, and successfully uncover active episodes of induced earthquakes, indicating the method’s potential as a useful tool for exploring primary earthquake occurrence patterns, especially in less-studied regions with sparse station coverage.