SES-07. Comparison of Neural Network Architectures for Seismic Arrival Picks In Time-Series Waveform Data
Abstract
Various neural network architectures have been proposed through the literature to
model anomalous signals in time-series data, including long short term memory networks (LSTMs), autoencoders and convolutional neural networks (CNNs). We compare performance of seismic arrival picks between LSTM and CNN architectures using 1D waveform data as input. While LSTMs are the more commonly used method for time-series modeling, CNNs have shown, when properly implemented, to outperform LSTMs in accuracy and efficiency. Generally, CNNs have shown to have lower computational costs which would benefit near-real-time seismic arrival picking models used within seismic networks. We propose the exploration of a novel CNN architecture to ingest 1D waveform data that will leverage CNN’s inductive bias with local spatial invariance to outperform the LSTM architecture. Additionally, we will perform a trade-off study of accuracy versus computational cost between a lightweight 1D waveform input and a more robust time-frequency spectrogram input within a CNN architecture.