Integrating Target Variable Selection Into Environmental Machine Learning for Surface Water Pollution Management

Target variable selection is often predefined in machine learning (ML)-based surface water pollution prediction without systematically considering dataset characteristics. This study developed a target variable selection unit (TVSU) integrating Gini feature importance and model pre-evaluation for same-period, cross-station, and cross-variable reconstruction of surface water quality and associated pollution-management decision support. Results showed that the scenario adhering to TVSU achieved the best predictive performance in both shallow and deep learning models (nRMSE values of 0.418 and 0.295/relative MAE values of 36.4% and 30.4%), whereas scenarios disregarding TVSU exhibited substantially higher prediction errors. Post hoc interpretation analyses further demonstrated that TVSU enabled clearer identification of key pollution-related variables. The proposed approach also provided flexibility for evaluating different pollution indicators and monitoring stations under multiple management scenarios. From a practical perspective, the workflow can help researchers and monitoring agencies prioritize candidate water-quality variables and monitoring stations before model development, thereby supporting more transparent and management-relevant environmental ML design. Overall, this study highlights the importance of systematic target variable selection for improving the reliability and applicability of environmental ML models in water quality management.