Designing Convolutional Neural Network Pipeline for Near‐Fault Earthquake Catalog Extension Using Single‐Station Waveforms
Josipa Majstorović, Sophie Giffard‐Roisin, Piero Poli
Abstract
Abstract In this study, we developed an end‐to‐end two‐stage pipeline using 1D convolutional neural networks (CNNs) to detect, localize, and characterize earthquakes from single‐station three‐component waveforms. We are presenting here the insights of what makes the difference in developing a deep learning algorithm by performing an extensive hyperparameter grid search for model training, tackling the question of the optimal number of classes, the importance of training data sets as well as the CNN architecture design in terms of optimal length of the CNN model. Moreover, our pipeline is robust and does not need any preprocessing of the seismograms (e.g., filtering) or any prior knowledge of the region. Training, validation, and evaluation of the CNN models is performed on data recorded at the AQU station placed in the city of L’Aquila in the Abruzzo region (Central Italy). Before MW 6.3 2009 L’Aquila earthquake that occurred near the city of L’Aquila the local catalog of events is sparse. Therefore, we provide a methodological pipeline on how one can extend near‐fault local catalog of earthquakes by applying our two‐stage pipeline on unseen continuous data. Our results show that we are able to design a CNN model that is detecting the earthquake events among random noise waveforms with 97% accuracy (first stage: detection). Additionally, we are able to determine the events that are close to the station (<10 km) with a 94% accuracy as well as identify their belonging to four magnitude classes with a 68% accuracy (second stage: characterization).