Anomaly detection on displacement rates and deformation pattern features using tree-based algorithm in Japan and Indonesia

Research on strain anomalies and large earthquakes based on temporal and spatial crustal activities has been rapidly growing due to data availability, especially in Japan and Indonesia. However, many research works used local-scale case studies that focused on a specific earthquake characteristic using knowledge-driven techniques, such as crustal deformation analysis. In this study, a data-driven-based analysis is used to detect anomalies using displacement rates and deformation pattern features extracted from daily global navigation satellite system (GNSS) data using a machine learning algorithm. The GNSS data with 188 and 1181 continuously operating reference stations from Indonesia and Japan, respectively, are used to identify the anomaly of recent major earthquakes in the last two decades. Feature displacement rates and deformation patterns are processed in several window times with 2560 experiment scenarios to produce the best detection using tree-based algorithms. Tree-based algorithms with a single estimator (decision tree), ensemble bagging (bagging, random forest and Extra Trees), and ensemble boosting (AdaBoost, gradient boosting, LGBM, and XGB) are applied in the study. The experiment test using real-time scenario GNSSdailydatareveals high F1-scores and accuracy for anomaly detection using slope windowing 365 and 730 days of 91-day displacement rates and then 7-day deformation pattern features in tree-based algorithms. The results show the potential for medium-term anomaly detection using GNSS data without the need for multiple vulnerability assessments.