Towards physically consistent prediction of groundwater salinization in coastal reclamation areas: A novel...

Groundwater salinization in coastal reclamation areas severely threatens regional water security. To address the limitation that machine learning models tend to overlook hydrogeochemical processes in black-box predictions, a PMF-weighted SHAP framework was proposed, which directly embeds source-apportionment information into feature selection. Qualitative evidence first indicated that trapped seawater in reclamation sediments, rather than active modern seawater intrusion, was the primary source of salinization. The PMF-derived contributions of trapped seawater were then integrated into SHAP as sample-specific physical weights, allowing samples with stronger trapped-seawater signatures to exert greater influence on feature ranking. Finally, following model benchmarking, XGBoost was selected as the base learner to predict the Groundwater Quality Index (GQI) using 407 groundwater samples, and three feature-input strategies were compared. The results show that compared to the conventional SHAP model, the PMF-weighted framework raised the importance ranking of land use and suppressed the spuriously high importance of tidal level, thereby aligning feature selection with the trapped-seawater-dominated salinization mechanism and increasing the test-set R² from 0.755 to 0.887. Stratified SHAP importance was further found to be highly consistent with the physical zonation revealed by PMF: irrigation predominantly affects the shallow zone (0-200 cm) in the paddy field, whereas trapped seawater dominates the deeper layer (300-400 cm) in both dryland and paddy field. This study demonstrates that embedding physical mechanisms into feature selection is an effective approach for constructing highly reliable predictive models in hydrology.