Earthquake Prediction using Long Short Term Memory on Spatio-Temporally Segmented Data

This paper describes a machine learning model for predicting earthquakes on the basis of past earthquake data. In particular, this study uses the Long Short Term Memory (LSTM) model, a neural network model designed to operate on time-series data with long-term dependencies. Here, the instrumental earthquake data is considered from three selected locations in Indonesia. First, the dataset is pre-processed by segmenting it into time intervals and space grids. The multi-dimensional time-series data is then fed into the network to output the probability of an earthquake in the next interval. This method was originally introduced by Wang et al. [1] and achieved an accuracy close to 85% on a dataset from Mainland China (1966–2016). To the best of our knowledge, no subsequent works have attempted to reproduce their results on different datasets, or introduce enhancements. This research work has implemented the same model on three different datasets. Further, the softmax activation function is replaced with the sigmoid activation function. This ensures that the probability values of earthquakes occurring in the segmented grids are independent of each other and are not rendered mutually exhaustive or exclusive events. Finally, a failure mode of this model is mentioned by showing that it performs poorly to predict large earthquakes.