Mendeley  

Zotero  

Pirizadeh M, Pirizadeh M. Artificial intelligence applications in analyzing seismological data (Case study: Precursors data). Disaster Prev. Manag. Know. 2021; 11 (3) :299-309
URL: http://dpmk.ir/article-1-430-en.html

Background and objective: Systematic recording of earthquake data in the last century has developed new approaches to using innovative methods to maintain, process, and analyze this type of data in order to organize and classify them for seismic risk management purposes. Among these approaches is the use of artificial intelligence technology to discover intrinsic rules and interdependencies between data to classify seismic events or to predict continuous values of time series.
Method: In this research, the applications of various branches of artificial intelligence, especially machine learning and deep learning methods in the processing of earthquake-related data have been analyzed by considering the specific characteristics of this type of data, including the temporal relationships between samples and the highly skewed distribution of classes. Then, a new approach based on deep recurrent networks equipped with cost-sensitive loss function is proposed to model the relationship between seismic time series data and the probability of future earthquakes. In order to evaluate the performance of the proposed model, a case study on the classification of Peak Ground Acceleration (PGA) severity in the next time steps has been performed based on the time series related to the seismic precursor of abnormal activity of animals.
Findings: The use of deep recurrent networks due to their ability to memorize long-term temporal relationships among samples, was evaluated quite positively in the modeling of seismic time series. However, in seismic data classification problems, in which more serious earthquakes occur less frequently than in non-serious cases, the problem of class imbalance plays a prominent role which, if not properly controlled, can greatly affect the performance of models in favor of the majority class. So that models will tend to label all the samples as a majority one, while in such situations the correct identification of the minority class has the topmost priority.
Results: The results of the case study revealed that the deep learning based approach with an average balanced-accuracy of 81.2% and 59.3% on training and test data, respectively, has left a completely better performance compared to conventional recurrent neural networks in the classification of PGA values. This study also shows that by modifying the loss function of deep recurrent networks using the cost-sensitive approach, the challenge of seismic data class imbalance can be well controlled.