Spatio-Temporal Seismicity Prediction in Chile Using a Multi-Column ConvLSTM

Abstract

One way to characterize the seismicity in a given zone is through the study of the conditional intensity function of the ETAS model (Epidemic Type Aftershock Sequence) which represents the average number of seismic events greater than given magnitude. Being Chile one of the most seismic country in the world, it is very important to predict where the seismic events will happen with more frequency. In this work we propose a parallel neural network based on the Convolutional Network (CNN) and the Long Short Term Memory (LSTM) network, called Multi-Culumn ConvLSTM, using the accumulated crustal velocity and the intensity data as input for predicting the daily mean number of seismic events in Chile with magnitude greater than a given value. For the application, the central zone of Chile between the regions of Coquimbo and Araucanía, in the period from 2010 to 2017 was considered. At the spatial level, each region was partitioned considering a $20\times 20$ dimension grid, while at the temporal level, input data from the last 20 days were used to predict the mean number of seismic events for the following day. The experiments showed that the Multi-column ConvLSTM network obtained the best results in the test set with an average coefficient of determination of 0.81.

Keywords

• Deep learning
• ETAS model
• prediction
• seismic events