Construction of Long-Term Seismic Catalog with Deep Learning: A Workflow
for Localized Self-Attention RNN (LoSAR)
Abstract
Characterizing fault behaviors prior to large earthquakes through
long-term seismicity is crucial for seismic hazard assessment, yet
constructing high-resolution catalogs over extended periods poses
significant challenges. This study introduces LoSAR, a novel deep
learning-driven workflow that enhances phase picking by Localizing a
Self-Attention Recurrent neural network with local data, addressing the
generalization problem common in data-driven approaches. We apply LoSAR
to two distinct regions that are both featured by recent large
earthquakes: (1) preseismic period of the Ridgecrest-Coso region
(2008-2019), and (2) pre-postseismic period of the East Anatolian Fault
Zone (EAFZ, 2020-2023/04). Through detailed comparisons, we demonstrate
that LoSAR offers slightly higher detection completeness than the QTM
matched filter catalog, while boosts an over 100 times faster processing
and a superior temporal stability, avoiding low-magnitude gaps during
background periods. Against PhaseNet and GaMMA, two established phase
picker and associator, LoSAR proves more scalable and generalizable,
achieving roughly 2.5 times more event detections in the EAFZ case,
along with a ~7 times higher phase association rate. By
leveraging the two enhanced catalogs and b-value analysis, we gain
insights into the preseismic fault behaviors: (1) The Ridgecrest faults
are characterized by sparse and distributed seismicity across a band of
~20 km, revealing multiple orthogonal preexisting
faults; coupled with a low b-value that signifies this area as a
persistent asperity; (2) The Erkenek-Pütürge segment of EAFZ exhibits
complex fault geometry that forms a persistent rupture barrier, which
consists of a hidden conjugate fault system that presents as a
~10-km wide fault zone.