Image classification using ConvNets: a reliable processing tool for fibre optic seismology

Matthaiou, Ioannis, Masoudi, Ali, Araki, Eiichiro, Kodaira, Shuichi, Modafferi, Stefano and Brambilla, Gilberto (2023) Image classification using ConvNets: a reliable processing tool for fibre optic seismology. Geophysical Journal International. (Submitted)

Abstract

Distributed acoustic sensing (DAS) is gradually being adopted as a standard tool to monitor and identify geophysical phenomena. In seismology, many researchers have been using DAS, instead of conventional sensing systems, for continuous long-term monitoring of natural phenomena over remote and harsh environments. Coupled with the ease of deploying DAS in an existing fibre optic cable that spans tens or hundreds of kilometres, seismologists are now in a better position to understand and analyse these naturally-occurring physical processes. This field of study is currently known in the literature as fibre optic seismology, and is commonly associated with large volumes of generated data, i.e. in the form of dynamic strain. A crucial aspect for the success of fibre optic seismology is the development of efficient and automated algorithms that can best make use of the large data volumes generated by a typical DAS device. In this study, a classifier is presented, which can be used to assign class-conditional probabilities for three different types of DAS strain measurements: geophysical, non-geophysical and background noise. These measurements were obtained using a 50 km-long optical fibre cable, which was deployed on the ocean floor close to the Nankai subduction zone, off the shore of Cape Muroto in Japan. Within this seismically active zone, where slow-slip events and very low-frequency earthquakes are common, a range of different hydroacoustic signals were identified and manually labeled by expert seismologists. The classifier, based on Convolutional Neural Networks (ConvNets), was then trained with a rather small set of experimental DAS observations (plots of time vs distance vs strain). In order to ensure good model generalisation capability the ConvNets' hyperparameters (including its architecture) were chosen using Bayesian Optimisation. The best model that was identified from the analysis yielded a 92 % overall classification accuracy on a dataset of 100 unseen DAS observations. Our results demonstrate the effectiveness and robustness of the proposed supervised learning model, for classifying DAS observations from multiple sources of hydroacoustic energy.

This record has no associated files available for download.

Contributors

Author: Ioannis Matthaiou

Author: Ali Masoudi

Author: Stefano Modafferi

Author: Gilberto Brambilla

Download statistics