Combining CNN and RNN in seismic phase picking

Abstract:

Recent expansion of seismic data and computing resources enables flourishing applications of deep learning in seismology (e.g., earthquake detection, phase picking, and phase association). Successful deep learning applications require an adequate amount of training data. The question, then, arises: how much data is enough? Here, we attempt to understand the data volume required for phase picking using the convolutional neural network (CNN). We adopt the labeled seismic data recorded by the Southern California Seismic Network (SCSN) compiled by Ross et al., 2018. The database is composed of 4.5 million 6-sec three‐component waveforms of P, S and noise earthquakes (1.5 million per category). We train a CNN classifier composed of 4 convolutional layers and 2 fully connected layers with only 1.4% of the dataset (64k records in each category). Our classifier only uses 24% of parameters in Ross et al., model, but has similar performances (96.5% accuracy). We consider 64k records are enough to extract the features of P and S phase windows.

We then test the generalization ability of both models, by adopting the STanford Earthquake Dataset (STEAD) with seismograms recorded by 2,613 globally distributed seismic stations. It contains 1 million 60-sec three-component waveforms labeled with P/S arrival time. By utilizing sliding windows, previously trained CNN models can output a time series of P/S/noise probability. If we set the threshold of 0.95 to pick phases, the precision and recall of both our and Ross et al., 2018 models are around 77% and 89%, respectively.

To reduce false/missing detections, we build a Long Short-Term Memory (LSTM) Recurrent Neural Network (RNN) on top of the previous trained CNN model. The CNN model extracts features from raw seismograms , while the RNN model recognizes and takes advantage of the time-related context. The input of RNN is the time series of predicted P/S/noise probability by CNN model while the output (target) is the converted masks for human labeled P/S arrivals (Zhu and Beroza, 2019). By including the additional RNN, we increase the precision and recall to 95% and 93%, respectively.