Diagnostic Factors for Earthquake Precursor from Earthquake Catalogs with Machine Learning-XGBoost Classifier

Precursors of large earthquakes are useful indicators for predicting future ruptures and their probability. Analyzing earthquake catalogs that contain occurrence time, epicentral locations, and earthquake magnitude is a fundamental tool to diagnose this precursor. Here, we trained a tree-based eXtreme Gradient Boost (XGBoost) Classifier with multi-featured datasets obtained from GCMT and ISC-GEM catalogs to recognize earthquake precursors for magnitude above magnitude 6.5. Training and validating datasets are composed of eighty augmented features from an earthquake catalog information (e.g., time, latitude, longitude, depth, and magnitude) in specific spatiotemporal windows. The label is assigned to each window according to the occurrence of events above magnitude 6.5 within three months. The results of five-fold cross-validation show that the accuracies of five validation sets are 0.4-0.8 with ~0.01 deviation. Furthermore, we adopted earthquake catalogs from other sources (e.g., JMA, USGS, and SCEC) to verify the predictive power of our model to untrained data. Although these values are lower than the validation results, model performance is enough to recognize precursory phases. We take advantage of tree-based machine learning algorithms to calculate feature importance explaining the contribution of each feature for the learning procedure. The two most important features in the window are i) the maximum magnitude of events and ii) interval time to the following events. These two features support previous studies suggesting the magnitude of each event in earthquake signal and increasing earthquake frequency before mainshock is a precursory indicator of large earthquakes. The classifiers with the most ten important features show higher accuracy by ~0.1-0.2. We suggest that our XGBoost Classifier approach may help to analyze the earthquake precursors, ultimately leading to earthquake hazard mitigation.