Enhanced Prediction of Persistent Earthquake-Induced Groundwater Level Changes with Advanced Feature Engineering and Machine Learning

Forecasting groundwater level fluctuations induced by seismic activity presents a considerable challenge due to the inherent complexity and pronounced non-linearity of the underlying processes. The limited availability of predictive variables further complicates this task, with key factors such as seismic shaking intensity, geological characteristics of dams, and shear wave velocity serving as primary indicators. To address the scarcity of predictive features and the intricate non-linear dependencies between input variables and groundwater level responses, we introduce an innovative fusion of feature engineering and machine learning. Our methodology is applied to a comprehensive regional-scale, multi-site, multi-earthquake dataset from New Zealand aquifers. Utilizing a filter-based supervised feature selection technique, we extract novel feature sets with strong correlations to groundwater level dynamics. Subsequently, we develop a random forest classification model to predict earthquake-induced groundwater level changes. The proposed approach significantly enhances both predictive accuracy and interpretability compared to conventional probabilistic models, offering a robust framework for improved seismic hydrogeological forecasting.