Büyük depremleri takip eden artçı şok örüntülerinin derin öğrenme ile modellenmesi

This thesis presents data-driven modeling strategies for predicting the spatial distribution, number, and magnitude of aftershocks following major earthquakes. To overcome the limitations of classical physics-based approaches, alternative model structures were developed for both classification and regression tasks using machine learning methods such as artificial neural networks, the XGBoost algorithm, and kernel-based probabilistic modeling. The study systematically evaluated various input combinations—including static components derived from stress tensors, neighborhood information, gradient-based derivative quantities, spatial coordinates, and distance to the epicenter—and additionally applied oversampling and undersampling strategies to mitigate class imbalance, while employing kernel-based continuous labeling to model the spatial continuity of aftershock probabilities. Through this comprehensive experimental analysis, the prediction performances of aftershock occurrence, count, and magnitude were compared across different data structures and modeling strategies, demonstrating the potential of stress-based machine learning models in aftershock forecasting. The findings indicate that kernel-based probabilistic models provide a more realistic and spatially coherent representation of aftershock regions, XGBoost enhances discrimination capability in classification tasks, and multilayer neural networks yield more balanced performance in numerical regression tasks. Overall, the results show that static features derived from stress tensors can successfully capture large-scale spatial patterns of aftershock behavior, yet modeling fine-scale variability will require richer temporal and geological datasets in future research.