<![CDATA[East Java is a seismically active region where short-term earthquake forecasting remains a critical yet challenging endeavor. While deterministic prediction is inherently unfeasible, probabilistic modeling offers a practical pathway for risk mitigation. This study develops a 30-day forward-window probabilistic forecasting model for M≥5.0 earthquakes in East Java using a Long Short-Term Memory (LSTM) network framed as a binary classification task. The model is trained on 25 years of seismic data (2001–2025) from BMKG Stasiun Geofisika Pasuruan. Twenty-five seismic features were rigorously selected through correlation analysis and data-leakage prevention protocols, while class imbalance was mitigated using adaptive loss weighting. The LSTM architecture was systematically optimized via sequential hyperparameter tuning and robust validation strategies. On a hold-out test set, the model achieved an AUC-ROC of 0.752, F1-score of 0.484, and recall of 0.673, indicating the model's capacity to detect impending seismic events with reasonable sensitivity. These results confirm that deep learning can effectively capture non-linear temporal patterns in seismic sequences. The primary contribution of this work is a validated, operationally ready probabilistic forecasting framework that can be integrated into regional earthquake monitoring systems, providing actionable lead time for disaster preparedness in East Java.]]>
Copyrights © 2026
