Loading...
Signal detection with neural networks in dark fiber optic seismic data: a case study from Istanbul to detect local seismicity
Hoyle, Austin Marino
Hoyle, Austin Marino
Citations
Altmetric:
Advisor
Editor
Date
Date Issued
2024
Date Submitted
Collections
Research Projects
Organizational Units
Journal Issue
Embargo Expires
Abstract
Distributed Acoustic Sensing (DAS) provides continuous and high-resolution spatial monitoring capabilities along the entire length of the optical fibers where the ground deformation, caused by seismic activity, is measured in strain or strain rate. Telecommunication cables, known as dark fibers, offer a spatially dense and cost-effective alternative for seismic data collection in urban areas. We demonstrated recording seismic signals with a telecommunication cable provided by the Istanbul Metropolitan Municipality along the Sea of Marmara on the Anatolian side of Istanbul from February to mid-March in 2023. The collected data involves various seismic signals, such as teleseismic and local earthquakes, including the East Anatolian earthquakes in February 2023 and their aftershocks, controlled explosions, traffic, and other cultural noise.
This research aims to demonstrate the detection of local seismic activity in and around Istanbul using DAS data. Monitoring local seismicity is essential to better understand the tectonics and seismic activity in earthquake-prone regions. However, it is challenging to detect and discriminate small earthquakes where the higher spatial resolution of DAS may be advantageous. To this end, we first visually inspect as many earthquakes (local, regional and teleseismic) listed in the earthquake catalogs recorded by the dark fiber that we could find and compare our observations to data recorded by nearby strong-motion seismometers. Our goal is to detect and explore the local events, through which we will use neural networks to do so. Since we have a limited data set from Istanbul, which makes the training challenging, we explore generating quasi-artificial data to build our training set. We do this with the intention of training our neural network that when allowed to process live DAS data in a sliding window fashion can flag event detection for researchers.
Our neural network performs with 98% - 100% success (depending on required threshold value to flag a seismic event) for both the quasi-artificial training set and a reserved dataset of real DAS data containing earthquake signals. We then let the neural network scan 38 days of nearly continuous data from the project in Istanbul in a sliding window fashion, analyzing nine-second windows at a time. Over the 38 days, it detected 46 seismic events. Of these 46, we have located the causing earthquake by searching the United States Geological Survey (USGS), Kandilli Observatory and Earthquake Research Institute (KOERI), and Disaster and Emergency Management Authority (AFAD) catalogs for 31 of them. Of the remaining 15, we were unable to confirm a cataloged event for them. This indicates they are either extremely large active sources, or may be previously unaccounted for/undocumented earthquakes.
Associated Publications
Rights
Copyright of the original work is retained by the author.