Artificial Intelligence–Driven Analysis of Seismic Events: Severity Classification and Epicenter Modelling in the Middle East

Abstract

It proposes a data-driven dual-task framework for seismic event analysis which combines both machine learning and deep learning approaches for classifying the severity level of earthquakes and predicting their epicenter coordinates. The dataset drawn from the Seismological Observatory in Jordan contained geophysical, geographical, as well as temporal features and went through preprocessing involving cleansing of data, normalization, as well as augmentation with synthetic Gaussian noise. These operations enlarged dataset to 10,000 rows and at same time more representative. In addition, earthquakes magnitudes were grouped into low, intermediate and high severity levels. Ensemble models (XGBoost, LightGBM, CatBoost and Gradient Boosting) with model training, performance validation for precision, recall F1 measure & ROC analysis were performed. The models showed high precision and best performance was achieved for Gradient Boosting with 99.67%. Furthermore, for the regression step, stacking ensemble of Multilayer Perceptron, Random Forest and XGBoost were applied for predicting epicenter coordinates. The model was highly predictive with respect to latitude (R^2 = 0.96; MSE = 0.70) and the same held for longitude, although it was systematically, but moderately less accurate (R^2 = 0.90; MSE = 4.39). Furthermore, confidence in results was provided through visualizations using residual plots, which also demonstrated practicality and usability of the developed method with interactive maps. It shows that our framework can deal with difficult seismic description and geospatial prediction problems by applying ensemble learning. It provides a basis to concrete the early warning and seismic hazard assessment system in middle east by improving its ability to reach very high intensity class determination and epicentral location approximation.