Combining CNN and RNN in Seismic Phase Picking

Tian Feng, & Lingsen Meng

Published August 15, 2020, SCEC Contribution #10727, 2020 SCEC Annual Meeting Poster #194

The recent expansion of seismic data and computing resources enables flourishing applications of deep learning in seismology. Many studies aim at automatically picking P and S arrivals, especially interested in microseismicity buried under noises. Dozens of deep-learning-based models prove to be efficient in detecting phases of local events (<300km). Most of them take seismograms/spectrograms as input data (X) and output the probability of P/S/background phases (Y).

Two different strategies developed when choosing the length of input data (X). The first strategy uses long windows of seismograms/spectrograms as input (>= 30sec), which contains the whole earthquake events. The corresponding model is sequence to sequence (N to N) type, in which the output is the probability of P/S/Background as a function of time (3 component sequence). The other strategy uses a fragment of seismograms/spectrograms as input (~ 4 sec), which only contains P or S or background signals. The P or S phases are at the center of the windows. The corresponding model is a sequence to one (N to 1) type, in which the output is the probability of P/S/background (a 3-component vector). When sliding this classifier on the long record of seismograms, it can generalize a sequence of the probability of P/S/Background.

Our model takes advantage of two strategies (small model and direct loss function). First, we train a CNN model (N to 1) to classify a small window segment belonging to P/S/background noise, composed of five convolutional layers and two fully connected layers. Instead of outputting the final layer (3 features/probabilities), we output the second last layer (128 features). We then train a Long Short-Term Memory (LSTM) Recurrent Neural Network (RNN) model on top of the CNN classifier. The input of RNN is a sequence of 128 features extracted by the last second layer of CNN, and output is the sequence of the probability of P/S/Background (N to N). The CNN model extracts features related to P and S phases from raw seismograms, while the RNN model recognizes and takes advantage of the time-related context. By including the additional RNN model, we increase the precision of P and S arrivals dramatically. Compared to the model trained directly with the N to N strategy (PhaseNet, Zhu et al., 2019) and N to 1 strategy (GPD, Ross et al., 2018), our hybrid model achieves better performance with fewer parameters.

Feng, T., & Meng, L. (2020, 08). Combining CNN and RNN in Seismic Phase Picking . Poster Presentation at 2020 SCEC Annual Meeting.

Related Projects & Working Groups
Computational Science (CS)