Development of a rapid and cost-effective groundwater quality assessment model based on hybrid ensemble learning

Assessing groundwater quality and health risks using machine learning is receiving widespread concern. However, assessment accuracy and cost-effectiveness are key factors in determining the model implementation. Therefore, the main purpose of this study is to develop a convenient, low-cost, and accurate hybrid ensemble model to predict water quality index (WQI) and hazard index (HI). Firstly, Pearson correlation matrix and ‘SHAP’ value were compared to select the Optimum feature combination. Secondly, base learners were selected from 12 different machine learning candidates. And then select eXtreme Gradient Boosting (XGB) as meta learner to construct stacking and blending ensemble model. The prediction results of the base learners are averaged to obtain the prediction results of averaging ensemble model. Finally, evaluation matrix (R2 and RMSE), t-test and probabilistic forecasting were integrated to assess models’ performance. The results show TDS, HCO3–, Mg2+, SO42- is the best feature combination for WQI prediction, and Na+, Ca2+, Mg2+, HCO3– is the best feature combination for HI prediction. SHAP value perform better than Pearson correlation matrix in reducing the number of input variables and improving model accuracy. The accuracy of stacking ensemble model on test/validation sets (average R2 = 0.966/0.921 and 0.835/0.714 for WQI and HI respectively) significantly (p < 0.05) higher than the other models. The Stacking ensemble model developed in this study provides supports for governments to assess groundwater quality and formulate rational policies. Meanwhile, the integration of evaluation metrics and statistical analysis also offers new ideas for model evaluation in the environmental field.