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

Feng CaiYifan LiuSheng LinWeiliang LiaoBeihan Jiang^*

• Fuzhou University, Fuzhou, China

Estuaries are dynamic hydrodynamic–biogeochemical interfaces where riverine and marine processes converge. Water quality is highly sensitive, functioning as an integrated response to both indirect and direct influences of meteorological variability and human disturbances modulated by tidal dynamics. Accurate of prediction of water quality in these environments is essential for maintaining ecosystem stability and reducing ecological risks. However, existing predictive approaches often suffer from incomplete monitoring data and limited capacity for multi-indicator modeling, thereby constraining either their accuracy or timeliness. This study proposed an enhanced Deep Forest–XGBoost framework (EDF-XGB) driven by high-resolution meteorological inputs. To enable adaptive parameter tuning, the global search whale optimization algorithm (GS-WOA) was incorporated along with a hierarchical feature selection procedure based on feature importance and a dynamic weighting mechanism that reflects sample difficulty. The proposed method was validated using a case study of the Min River Estuary. The results indicated that the EDF-XGB model achieved high predictive accuracy for stable indicators such as pH, total nitrogen (TN), and dissolved oxygen (DO) (R² > 0.90), and demonstrated evident advantages for more variable parameters, includin ammonia-nitrogen (NH₃-N) and the Permanganate Index (CODMn). Interpretability analysis via SHapley Additive exPlanations (SHAP) identified water temperature(WT), surface temperature(ST), and relative humidity (RH) as the dominant drivers of water-quality dynamics, consistent with the physical and chemical processes governing estuarine hydrochemistry. Regional generalization tests indicated strong predictive performance in the upstream non-tidal sections, whereas the accuracy decreased in the downstream tidal reaches influenced by hydrodynamic and anthropogenic factors, highlighting the need to incorporate hydrodynamic and human activity descriptors into future model development. The interpretable, data-driven, multi-indicator framework presented in this study can provide a scientific basis for real-time prediction and ecological risk warning in estuarine systems, contributing to enhanced meteorological resilience and sustainable management of vulnerable coastal environments.