Next Generation Earthquake Monitoring: Harnessing Deep Learning for Enhanced Seismic Phase Detection and Association
Author(s)
Chuang, Yuling Lindsay
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
As seismology enters the era of big data, the exponential growth in data volume and processing needs surpasses the capacity of traditional seismic monitoring workflows. The recent success of machine learning applications across various scientific domains has made a paradigm shift in image processing and simple task automation. Within this context, this thesis presents a modern earthquake monitoring workflow with deep learning integrated into different fronts. In the first study, I utilized a deep learning phase picker - EQTransformer and a template matching method for foreshock discovery. I performed a detailed study of the foreshock sequence preceding the 2010 magnitude 6.7 Yushu, Qinghai earthquake in the Tibetan plateau and successfully identified 120 foreshocks with magnitude ranging from -0.7 to 1.6. The foreshock sequence started with an magnitude 4.6 foreshock that occurred approximately 2 hours before at a fault plane roughly perpendicular to the mainshock rupture zone. The observations suggest that extensional step-overs and conjugate faults along major strike-slip faults play an important role in generating short-term foreshock sequences. In the second and third studies, I introduced a novel phase association and location framework tailored for a global-scale seismic monitoring network. The global seismic phase association remains a challenging task due to several factors, such as an inhomogeneous sparse seismic network, the high volume of phase arrivals (comprising both true and false picks), and the large solution space inherent for the global scale. I crafted a framework to tackle these challenges by combining an ensemble deep learning locator, advanced sampling strategies, beam search, and an OcTree grid search algorithm. Through comprehensive evaluations with synthetic and real-world datasets, I demonstrated the framework's effectiveness in associating seismic phases, even in scenarios with multiple events, noise, and overlapping events. During a 9-day trial period in May 2010, this framework recovered up to 93 \% of the events cataloged in the analyst-curated Unconstrained Global Event Bulletin (UGEB) catalog when applied to the phase arrival dataset at the International Data Centre (IDC), while effectively handling up to 88 \% of false picks, despite only using P-waves. Finally, I presented an integration of full waveforms and velocity models through an auto-encoder network and an Eikonet-style deep-learning surrogate model. This work contributes to the modern earthquake monitoring workflow by leveraging deep learning across various aspects of seismic research in the era of big data.
Sponsor
Date
2023-12-11
Extent
Resource Type
Text
Resource Subtype
Dissertation