Automated Artificial Intelligence Approach for Reliable Earthquake Precursors Detection and Classification Using Hybrid Deep Learning Models

Abstract

The precise detection of earthquake precursors before P-wave arrival is essential for advancing real-time seismic monitoring and refining earthquake source parameter estimation. Traditional P-wave detection methods frequently encounter ambiguity stemming from the simultaneous presence of instrumental artifacts and natural groundorigin precursory signals, leading to misclassification and timing discrepancies. An automated artificial intelligence system has been designed to reliably classify precursory signal patterns, thereby overcoming these constraints. The analysis utilizes a synthetically constructed dataset that mirrors the properties of precursor patterns seen in seismic recordings obtained at various sampling rates. Dataset exploration and visualization are conducted by trend strength distributions and correlation analysis, succeeded by preprocessing and an 80:20 train-test division. Various classification models are employed, including Logistic Regression, K-Nearest Neighbors, Support Vector Machine, Decision Tree, Random Forest, Naive Bayes, XGBoost, Voting Classifier, and convolutional neural networks, utilizing diverse data splits, in addition to sophisticated hybrid Stacking and BiLSTM architectures. The evaluation of model performance is conducted by accuracy, precision, recall, F1-score, and ROC-AUC metrics, supplemented by comparison performance graphs. The BiLSTM model demonstrates superior performance across all assessed methodologies, with an accuracy of 99.8% and an almost flawless ROC-AUC, markedly exceeding the results of baseline and ensemble models. Explainable AI methodologies utilizing LIME and SHAP facilitate the interpretation of feature contributions, while their integration with a Flask-based framework allows for real-time user input, preprocessing, prediction, and result visualization, thus enhancing automated precursor classification and P-wave arrival determination