Real‐Time Earthquake Detection and Magnitude Estimation Using Vision Transformer

Abstract

Earthquake magnitude estimation is one of the fundamental processes for earthquake monitoring systems. The earthquake magnitude denotes the energy released during the shock, which is one of the main concerns for scientific research and the public. Richter magnitude (local magnitude) is one of the most widely used scales to inform the public of the earthquake size (Richter, 1935). Empirically, Richter magnitude (M L ) is obtained for every single station (each trace) by estimating the maximum amplitude among certain periods and takes the average magnitude over multiple stations (Havskov & Ottemoller, 2010). However, measuring the maximum amplitude of the seismic stations can be affected by unwanted signals of, for example, industrial activities, interfering noise, random noise, or other sources of interference, which leads to inaccurate magnitude estimation (Yamada & Mori, 2009). There are two strategies for magnitude estimation, that is, on-site (single station) approach (Wu et al., 2006) and network approach (Kuang et al., 2021). One single station is used in the on-site approach,

Microseismic Event Classification With Time-, Frequency-, and Wavelet-Domain Convolutional Neural Networks

“…The CNN model of [25] reports classification results that are among the best in the literature, without complex feature engineering. Although using more complex input features and network structures, such as [26], [13], [22], [24], [23], can lead to high performance, the complex preprocessing steps and deep models are computationally demanding. In addition, sequence-to-sequence learning requires clearly labelled start and end times of each event.…”

“…The deep network structure consists of an encoder that converts the raw input signal into features through 1-D convolution, max-pooling, residual convolution, and LSTM layers, and 3 separate decoders. In [23], the authors propose a vision transformer (ViT)-based system for earthquake detection and its magnitude prediction. The system consists of two separate ViT networks: the first one detects earthquake events from the picked P-wave; the second network predicts the magnitude of the detected earthquakes.…”

Machine Learning Developments and Applications in Solid‐Earth Geosciences: Fad or Future?

“…In this special collection, multiple studies further improve the robustness and generalizability of ML‐based earthquake detection. The improvements are achieved by utilizing a vision transformer architecture (Saad et al., 2022), by designing cascaded neural networks (Majstorovic et al., 2021), by data augmentation (T. Wang et al., 2021) and transfer learning (Lapins et al., 2021), by transforming seismic data into the time‐frequency domain before detection (Saad et al., 2021), and by incorporating higher abstraction features and latent space information over the seismic array (Mosher & Audet, 2020; Z. Xiao et al., 2021; Feng et al., 2022). Baseline neural networks are trained using massive labeled datasets with several tens of thousands of data entries, while transfer learning reduces this requirement to a few thousand.…”

Automatic microseismic signal classification for mining safety monitoring using the WaveNet classifier

“…The short‐term average/long‐term average technique is a popular algorithm to detect microseismic events in continuously recorded signals (Allen, 1978). Recently, deep neural network–based techniques have also been actively developed (Chen et al., 2019; Mousavi et al., 2020; Saad & Chen, 2020; Saad et al., 2021, 2022; Zhang et al., 2020, 2021). However, we adopted a simple amplitude threshold‐based picking method to increase recording efficiency.…”