Machine Learning–based Earthquake Probability Prediction for Istanbul Province

Abstract

Earthquakes represent one of the most destructive natural hazards, causing severe social and economic losses, particularly in tectonically active regions. In this context, estimating the short-term probability of earthquake occurrence is of critical importance for disaster preparedness and risk mitigation. This study focuses on Istanbul province, located along the North Anatolian Fault Zone in the Marmara Region, which is considered one of the most seismically active areas in Türkiye. The dataset used in this study consists of 633 earthquake events recorded between 2003 and 2025, compiled from the Disaster and Emergency Management Authority (AFAD), United States Geological Survey (USGS), and SeismicPortal catalogs. Machine learning approaches were applied to estimate whether earthquakes with magnitudes of M ≥ 3.5 could occur within a seven-day forecasting horizon. Three models—Random Forest (RF), Support Vector Machines (SVM), and Long Short-Term Memory (LSTM)—were trained and evaluated using several performance metrics, including MAE, RMSE, and R². Due to the imbalanced nature of the dataset, classification-based metrics such as ROC-AUC and F1-score were also evaluated to provide a more reliable assessment of model performance. The comparative analysis shows that ensemble-based models provide more stable predictive performance in short-term seismic forecasting tasks. The results demonstrate that machine learning–based probabilistic modeling can provide meaningful insights for short-term earthquake risk assessment and may support decision-making processes in disaster management and early warning systems.