Application of an enhanced deep forest model driven by meteorological dates in water quality prediction: a case study of the Minjiang River Estuary, Southeastern China

Introduction Estuaries are dynamic hydrodynamic–biogeochemical interfaces where riverine and marine processes converge, and their water quality is highly sensitive to meteorological variability and human disturbances modulated by tidal dynamics. Accurate prediction of water quality in estuarine environments is essential for maintaining ecosystem stability and reducing ecological risks. However, existing prediction approaches are often limited by incomplete monitoring data and insufficient capability for multi-indicator modeling, which constrains their accuracy and timeliness. Methods This study proposes an enhanced Deep Forest–XGBoost framework (EDF-XGB) driven by high-resolution meteorological inputs for multi-indicator water quality prediction. A global search whale optimization algorithm (GS-WOA) was integrated for adaptive parameter tuning, together with a hierarchical feature selection strategy based on feature importance and a dynamic weighting mechanism to account for sample difficulty. The proposed model was evaluated through a case study in the Min River Estuary. Results The results demonstrate that the EDF-XGB model achieves high predictive accuracy for relatively stable water quality indicators, including pH, total nitrogen (TN), and dissolved oxygen (DO), with R² values exceeding 0.90. For more variable indicators, such as ammonia nitrogen (NH₃-N) and the permanganate index (CODMn), the proposed model shows clear performance advantages over conventional approaches. SHapley Additive exPlanations (SHAP) analysis reveals that water temperature (WT), surface temperature (ST), and relative humidity (RH) are the dominant drivers of water quality variability. Discussion Regional generalization experiments indicate strong predictive performance in upstream non-tidal sections, whereas prediction accuracy decreases in downstream tidal reaches affected by complex hydrodynamic conditions and anthropogenic activities. This suggests that incorporating hydrodynamic descriptors and human activity indicators could further improve model performance. Overall, the proposed interpretable, data-driven, multi-indicator framework provides a scientific basis for real-time water quality prediction and ecological risk early warning in estuarine systems, supporting improved meteorological resilience and sustainable management of vulnerable coastal environments.