Achieving a Comprehensive Microseismicity Catalog through a Deep-Learning-Based Workflow: Applications in the Central Ecuadorian Subduction Zone

Although seismological networks have densified along the Ecuadorian active margin since 2010, visual phase reading, ensuring high arrival times quality, is more and more time-consuming and becomes impossible to handle for the very large amount of recorded seismic traces, even when preprocessed with a detector. In this article, we calibrate a deep-learning-based automatized workflow to acquire accurate phase arrival times and build a reliable microseismicity catalog in the central Ecuadorian forearc. We reprocessed the dataset acquired through the OSISEC local onshore–offshore seismic network that was already used by Segovia et al. (2018) to produce a reference seismic database. We assess the precision of phase pickers EQTransformer and PhaseNet with respect to manual arrivals and evaluate the accuracy of hypocentral solutions located with NonLinLoc. Both the phase pickers read arrival times with a mean error for P waves lower than 0.05 s. They produce 2.7 additional S-labeled picks per event compared to the bulletins of references. Both detect a significant number of waves not related to seismicity. We select the PhaseNet workflow because of its ability to retrieve a higher number of reference picks with greater accuracy. The derived hypocentral solutions are also closer to the manual locations. We develop a procedure to automatically determine thresholds for location attributes to cull a reliable microseismicity catalog. We show that poorly controlled detection combined with effective cleaning of the catalog is a better strategy than highly controlled detection to produce comprehensive microseismicity catalogs. Application of this technique to two seismic networks in Ecuador produces a noise-free image of seismicity and retrieves up to twice as many microearthquakes than reference studies.