We use cookies to ensure that we give you the best experience on our website. By continuing to browse this repository, you give consent for essential cookies to be used. You can read more about our Privacy and Cookie Policy.

Durham e-Theses
You are in:

Artificial intelligence to detect and forecast earthquakes

ONG, VEDA,LYE,SIM (2021) Artificial intelligence to detect and forecast earthquakes. Masters thesis, Durham University.

PDF - Accepted Version


Precursors to large earthquakes have been widely but not systematically identified. The ability of deep neural networks to solve complex tasks that involve generalisations makes them highly suited to earthquake and precursor detection. Large moment magnitude (Mw) earthquakes and associated tsunamis can have a huge economic and social impact. Detecting precursors could significantly improve seismic hazard preparedness, particularly if precursors can assist, within a more general probabilistic forecasting framework, in reducing the uncertainty interval on expected earthquakes’ timing, location and Mw. Additionally, artificial intelligence has recently been used to improve the detection and location of smaller earthquakes, assisting in the completion and automation of seismic catalogues.

This paper is the first to present a deep learning-based solution for detecting and identifying short-term changes in the raw seismic signal, correlated to earthquake occurrence. Deep neural networks (DNNs) were employed to investigate the background seismic signal prior to 31 Mw >= 6 earthquakes in the Japan region. Instantaneous, precursor-related features (features correlated to the investigated earthquakes) were detected as opposed to predicting future values based on previously observed values in the case of time series forecasting. The network achieved a 98% train accuracy and a 96% test accuracy classifying noise unrelated to Mw >= 6 earthquakes from signal immediately prior to the investigated earthquakes. Additionally, the precursor-related features became increasingly systematic (more frequently detected prior to the investigated earthquakes) with earthquake proximity. Discriminative features appeared most dominant over a frequency range of ~ 0.1-0.9 Hz, coinciding with microseismic noise and recent observations of broadband slow earthquake signal (Masuda et al. 2020). In particular, frequencies of ~ 0.16 and ~ 0.21 Hz provided significant precursor-related information.

Deep learning successfully detected features of the seismic data correlated to earthquake occurrence. Developing a better understanding of the origin of the precursor-related features and their reliability is the next step towards establishing an earthquake forecasting system.

Item Type:Thesis (Masters)
Award:Master of Science
Faculty and Department:Faculty of Science > Earth Sciences, Department of
Thesis Date:2021
Copyright:Copyright of this thesis is held by the author
Deposited On:06 May 2021 11:13

Social bookmarking: del.icio.usConnoteaBibSonomyCiteULikeFacebookTwitter