Date of Award
2022
Degree Type
Dissertation
Degree Name
Doctor of Philosophy in Oceanography
Specialization
Marine Geology and Geophysics
Department
Oceanography
First Advisor
Meng (Matt) Wei
Abstract
Largest earthquakes with destructive tsunamic waves mostly occurred in the offshore subduction zone, causing massive fatalities and significant property losses. Due to the limitations of the seafloor geodesy, it is hard to know the stress status offshore, like the terrestrial geodesy. The shallow slow slip event (SSE) occurrence provides an approach to studying the shallow subduction zone stress status and investigating the area and size of the potential large earthquake and following tsunamis. SSE is a bridge linking the slip rates from aseismic creeping near the trench to the highly locked region in the seismogenic zone. Seafloor pressure measurement, as the high precision, low cost, and continuous vertical deformation records, is the most common way of studying offshore vertical deformation caused by geodetic movements. However, due to the long-term instrumental drift and considerable water movement noise in the data, detecting and measuring shallow SSEs from the seafloor pressure data is still very hard.
In manuscript one, we developed a machine learning detector to detect the slow slip event in seafloor pressure data. Because real seafloor pressure data is not this abundant, we first trained the machine learning detector using synthetic data, and then applied the well-trained detector to the real seafloor pressure data collected by the HOBITSS project in New Zealand. The trained model can successfully detect an SSE and the accuracy increases with SSE amplitude. The synthetic data test also shows that the machine learning model outperformed the traditional matched filter method. Our detector found five events in real pressure data in New Zealand between 2014-12015, two of which are confirmed by the onshore GPS records.
In manuscript two, we applied our machine learning detector to Alaska. The Southern Alaska subduction zone is a high seismic risk zone. Megathrust earthquakes and following devastating tsunami waves threaten south Alaska and the entire Canada and US west coast. We want to know the stress status at the shallow subduction zone in southern Alaska. In this study, we improved our previous machine learning detector by detecting both uplift and subsidence signals. We found four adjacent stations at 100-m water depth were uplifted, while four adjacent stations near the trench subsided in days 290-310 of 2018. This pattern is unlikely oceanographic in origin, based on an analysis of 10-year model output from an ocean circulation model (HYCOM). This pattern is consistent with a simulated ground deformation from a circular SSE on the subduction interface. We also investigated the daily seismicity and tremors using both onshore and offshore seismometers. We found that these seismic activities are closely related to our detected SSE: (1). A few tremors occurred before and near the SSE area. (2). The increased seismicity after SSE is located at the positive stress change area. Furthermore, our detected SSE is located 150-km northeast of the 2020 Mw 7.8 Alaska earthquake and the updip of the 2021 Mw 8.2 Chignik earthquake. The SSE has possibly increased the Coulomb stress in the area of these two large earthquakes.
The major problem for previous papers is correctly and broadly removing the oceanographic signals as much as possible. Therefore, in Chapter Three, we investigated the water column movement contribution to seafloor pressure by combining the machine learning method and ocean circulation model. We first applied the random forest method to study different features' importance in predicting seafloor pressure data, and then we used deep learning neural networks to better predict the results by incorporating the time information. This is an ongoing project. For future work, the well-trained model can be generalized using a small part of real data, and more available observations can be used to predict the real seafloor pressure.
Recommended Citation
He, Bing, "DETECTING SLOW SLIP EVENTS FROM SEAFLOOR PRESSURE DATA USING MACHINE LEARNING" (2022). Open Access Dissertations. Paper 1451.
https://digitalcommons.uri.edu/oa_diss/1451
Terms of Use
All rights reserved under copyright.