What Are More Important for Aftershock Spatial Distribution Prediction, Features, or Models? A Case Study in China
Zhao Sha,Wang Haiyan,Xue Yan,Wang Yilin,Li Shijian,Liu Jie,Pan Gang
DOI: https://doi.org/10.1007/s10950-021-10044-x
2021-01-01
Journal of Seismology
Abstract:Aftershocks can cause disasters again after mainshocks, which result in threat to life and economic loss. In order to avoid secondary disasters, it is necessary to predict whether aftershocks would happen in a given region. There have been studies using different features and methods to predict aftershocks spatial distribution. However, it is still unclear which are more important for aftershock prediction, input features or models; which type of features is more predictive for the prediction task. In this paper, we predict aftershock spatial distribution by combining different types of features and applying different machine learning methods. We introduce five different types of features and combine them together for prediction: the stress change sensors, their logarithmic values, the physical quantities, the magnitude of mainshocks, and the distance between the grid cell and the epicenter of mainshocks. We train different classifiers: Naive Bayes, Support Vector Machine,Gradient Boosting Decision Tree, k-Nearest Neighbors, Logistic regression, and DMAP (a Deep Neural Network model). Based on the 62,811 aftershocks of 171 distinct mainshocks in the past about 40 years in China, we conduct comprehensive experiments and analyses. We find that features play a more important role for this prediction task. Using the same feature type, different classifiers obtain quite similar performance. With different features, the same model performs differently. Taking the combined features as input, we achieve the state-of-the-art performance, with an AUC of 0.9530, about 4% higher than that of DeVries et al., showing the superiority of the combined features. Among all the features, adding the distance to the stress change sensors contributes the most to improve the prediction performance. In addition, it is found that the model prediction performance varies in terms of the time spans after mainshocks and the aftershock magnitudes.