Precursory Pattern Based Feature Extraction Techniques for Earthquake Prediction

Abstract

Earthquake prediction is an important and complex task in the real world. Although many data mining-based methods have been proposed to solve this problem, the prediction accuracy is still far from satisfactory due to the deficiency of feature extraction techniques. To this end, in this paper, we propose a precursory pattern-based feature extraction method to enhance the performance of earthquake prediction. Especially, the raw seismic data is firstly divided into fixed day time periods, and the magnitude of the largest earthquake in each fixed time period is labeled as the main shock. The precursory pattern is a part of the seismic sequence before the main shock, on which the existing mathematical statistic features can be directly generated as seismic indicators. Based on these precursory pattern-based features, a simple yet effective classification and regression tree algorithm is adopted to predict the label of the main shock in a pre-defined future time period. The experimental results on two historical earthquake records of the Changding-Garze and Wudu-Mabian seismic zones of China demonstrate the effectiveness of the proposed precursory pattern-based features with the selected CART algorithm for earthquake prediction.