Context-Aware CNN with Attention for Automatic Earthquake Precursor Detection

Abstract

Anomalous seismic, environmental and geophysical signals are apparent occurrences that may operate as earthquake precursors, warning people of approaching seismic events. The accurate identification of these precursors, and the separation between real signals from the ground and artifacts from the experiment, are crucial for accurate P-wave detection and determination of earthquake source parameters. A dataset of 429 signals was obtained from the New Abu Dabbab station at three different sampling rates (50, 100 and 200 samples/sec) and classified as ramping, non-ramping or mixed patterns. For the automatic categorization of P-wave data, a Convolutional Neural Network (CNN) model was tested, as well as conventional machine learning algorithms like Logistic Regression, K-Nearest Neighbors, Support Vector Machine, Decision Tree, Random Forest, and XGBoost. The CNN model achieves an accuracy of 97.98% which is higher than the other models. Adding attention mechanism increased the prediction accuracy to 99.81%, indicating the superior comprehension of context and feature optimization in sequential seismic data. The robustness of the deep learning-based method was demonstrated by a complete examination through accuracy, precision, recall, F1-score and ROC curve analysis. Our results demonstrate that attentionaugmented CNN models offer a reliable and efficient framework for real-time P-wave signal classification, hence reducing mistakes in earthquake precursor identification.