Predicting the Magnitude and Location of Earthquake in Sichuan-Yunnan Region via a Convolutional Neural Network

Weifeng Shan, Yuntian Teng, Shengfeng Zhang, Maofa Wang, & Guanze Yang

Published August 15, 2020, SCEC Contribution #10653, 2020 SCEC Annual Meeting Poster #094

Deep learning algorithms have been successfully applied in the fields of micro-earthquake detection, earthquake locating, earthquake early warning and so on, which is obviously that seismic waveform data is mainly used in this research. At the same time, many kinds of geophysical background field observation equipment have been carried out in China and output a massive observation data for many years. However, the monitoring data of parts of these equipment have obvious fluctuations before an earthquake and make the data quality not very well. Therefore, how to comprehensively use these massive seismic observation data to predict earthquakes is a significant topic.

This work takes the Sichuan-Yunnan region as the research region where strong earthquakes often occur, and proposes a short-term and imminent earthquake prediction model based on convolutional neural network (CNN). We construct sample data according to the following steps: 1) Use the epicenter location data of historical earthquakes in the Sichuan-Yunnan area to divide it into 6 prediction blocks B using the K-means++ algorithm; 2) Use 240 hourly observation data of 11 measurement items including the static water level, dynamic water level, deformation and geomagnetic data of the Tonghai and TonghaiGaoDa seismic stations from 00:00 on January 1, 2015 to 23:00 on December 31, 2018, to construct an input sample data, and then slide for 30 hours to construct a new sample. The data of 11 measurement items are input into the CNN network as 11 channels; 3) Label training samples. If the magnitude of the biggest earthquake event in one sample is less than 3, the magnitude label M is marked as 0, or is marked as 1. In order to use CNN to forecast the location and magnitude of earthquakes in the next 10 days, we transform label B and M into one label E, 0 is marked if there is not earthquake event occurred during the sample data period. PCA down-sampling method and cross entropy loss function are also used in our CNN earthquake prediction model, which contains 1 convolutional layer, 1 Batch Normalization layer, 1 maximum pooling layer and 2 fully connected layers. Experimental results show that the accuracy rate of the prediction model reaches 95.0% and the recall rate reaches 84.1%, which provides a novel idea for earthquake prediction research.

Key Words
machine learning, convolutional neural networks,clustering algorithm, earthquake forecasting

Shan, W., Teng, Y., Zhang, S., Wang, M., & Yang, G. (2020, 08). Predicting the Magnitude and Location of Earthquake in Sichuan-Yunnan Region via a Convolutional Neural Network. Poster Presentation at 2020 SCEC Annual Meeting.

Related Projects & Working Groups
Earthquake Forecasting and Predictability (EFP)