Comparative analysis of TGAN and other GAN models for synthetic earthquake data: a case study with data from Türkiye

Early detection of earthquakes is very important to prevent possible loss of life and injuries. Türkiye faces frequent earthquake disasters due to its geographical location. In order to predict the possible earthquake risk with artificial intelligence methods, data is required. This paper explores the potential of synthetic data generation, focusing on the data constraint in simulating earthquake data and further analyzing disaster scenarios. TGAN, CTGAN and CopulaGAN models are compared using real earthquake dataset obtained from Istanbul Metropolitan Municipality open data portal. The results show that the TGAN model achieves the highest performance in both statistical and structural metrics. TGAN produced results close to the real data in terms of mean (19.53 vs. 19.84) and cumulative total (27,269.58), and obtained the highest value (0.9022) in correlation analysis. Kolmogorov–Smirnov (KS) test and chi-squared (CS) test results showed that all models modeled discrete attributes better, while the logistic regression classifier TGAN performed moderately well in distinguishing real data from synthetic data. These findings reveal that the TGAN model is an effective tool in the synthetic generation of earthquake data and offers new perspectives in disaster management processes. As one of the first comprehensive comparisons of the potential of GAN models for synthetic generation of earthquake data, this study makes an innovative contribution to the literature in terms of both model selection guidelines and synthetic data applications.