Graph‐Partitioning Based Convolutional Neural Network for Earthquake Detection Using a Seismic Array

Abstract

Enormous volumes of continuous seismic data have been acquired from seismograph networks over the past decade, with these data sets consisting of observations from multiple seismic stations. Dense seismograph networks, such as the Japanese Metropolitan Seismic Observation network (MeSO-net) and the Southern California Seismic Network, monitor real-time seismicity and provide continuous waveforms from their respective network stations. Efficient and thorough analyses of these data sets should be of great benefit to seismology. The main objective of the present work, which represents a novel approach to and advance in seismic data analysis, is the development of an improved earthquake detection technique for these massive seismic network data sets.In recent years, deep neural networks have been attracting increasing interest as tools for analyzing such complex big data in many applied fields such as image processing (

An Early Warning System for Earthquake Prediction from Seismic Data Using Batch Normalized Graph Convolutional Neural Network with Attention Mechanism (BNGCNNATT)

“…Latitude, longitude, depth, and seismic magnitude are displayed on the graph. GNN predicts graph properties by collecting and analyzing node information [16,27].…”

“…Based on geophysical array data, a model using CNN and GNN to predict earthquakes was put out by Ref. [27]. This model is based on graph partitioning technique.…”

“…For earthquake prediction, a graph partitioning approach using CNN has been proposed in Ref. [27]. These methods only employed conventional graph theory rather than the true GNN technique.…”

MagInfoNet: Magnitude Estimation Using Seismic Information Augmentation and Graph Transformer

“…Besides, different from the recurrent neural networks (RNN), the processing ability for time series of CNNs is not very powerful (Karim et al, 2018). Nevertheless, some reliable approaches such as the combination with the graph theory (Yano et al, 2021) and the Transformer model (Xiao et al, 2021), can provide alternative architectures for a more effective extraction of the temporal characteristics of seismic waveforms.…”

Spatiotemporal Graph Convolutional Networks for Earthquake Source Characterization

“…However, global pooling may not sufficiently extract all useful information from multiple seismic stations, as the pooling layer is ideally applied after global features are obtained by feature fusion along the spatial dimension. Yano et al (2021) introduce a multistation technique in which edges are selected and held fixed for all inputs. While this model allows for more meaningful features to be constructed than in global pooling, station inputs are required to be fixed during training and implementation, introducing the same limitation inherent to CNNs.…”