|
Title |
Groundwater Level Prediction Modeling in Coastal Aquifers Using Tree-Based Ensemble Machine Learning and SHAP
|
|
Authors |
최용환(Yonghwan Choe) ; 정관호(Kwanho Jeong) |
|
DOI |
https://doi.org/10.15681/KSWE.2026.42.4.301 |
|
Keywords |
Coastal aquifer; Groundwater level prediction; Time-series forecasting; Hybrid feature selection; Tree-based ensemble machine learning |
|
Abstract |
Accurate prediction of groundwater levels (GWL) is essential for the sustainable management of coastal aquifers, which are particularly susceptible to saltwater intrusion and climate change. This study developed a comprehensive prediction framework for a coastal aquifer in Busan, South Korea, featuring a complex ria coastline. We utilized six tree-based machine learning models: Decision Tree, Random Forest, Gradient Boosted Trees, XGBoost, LightGBM, and AdaBoost. A hybrid feature selection approach, combining Pearson correlation analysis and recursive feature elimination, effectively identified five optimal predictors from twelve hydrometeorological and water quality variables. After hyperparameter tuning through Bayesian optimization, AdaBoost achieved the best predictive performance (RMSE=0.040, NSE=0.996, R²=0.998). Taylor diagram analysis showed that the optimized ensemble models accurately captured the complex variability of coastal GWL, achieving high correlation (R>0.99) and low error (RMSD<0.06). Additionally, SHAP (SHapley Additive exPlanations) analysis revealed a statistical masking effect, where the strong temporal persistence of lagged GWL (GWL at t-1) obscured the short-term influences of meteorological variables. These results underscore the effectiveness of tree-based ensemble models as real-time monitoring tools for predicting near-future GWL dynamics. To improve proactive disaster management, including drought and inundation forecasting, future research should focus on developing exogenous models that exclude lagged GWL, allowing for a clearer understanding of the independent effects of meteorological variables and extending the forecasting horizons.
|