Research on the Model-Feature System Matching Framework and Adaptive Feature Association for Accurate Estimation of DO in Nearshore Ports

Dissolved oxygen (DO) is a key indicator for assessing the ecological health of nearshore harbors and managing the aquatic environment. Its concentration dynamics are governed by the nonlinear coupling of multiple factors, and obtaining accurate, interpretable estimates remains a central challenge in water quality modeling. Current research on DO estimation generally suffers from a disconnect between feature selection and model building, neglecting the systemic compatibility between the two. This results in suboptimal model configurations with poor site adaptability, redundant features that compromise model accuracy, and weak explanations of physical processes, making it difficult to meet the demands for refined water quality early warning systems. To address this issue, this study developed a Model-Feature System Matching (MFSM) framework based on 35 years (1988–2023) of monthly water quality monitoring data from Victoria Harbor in Hong Kong. By comparing all possible combinations of eight feature importance algorithms and five machine learning models, the framework enables adaptive diagnosis and adaptation of optimal modeling schemes for DO estimation at specific monitoring sites within the harbor. The results indicate that the optimal configuration for Station VM1 is a Pearson feature selection-coupled MLP model, which achieves the highest estimation accuracy (RMSE = 0.1292 mg·L−1, R2 = 0.9904). The Random Forest model demonstrated the most significant improvement in accuracy after feature optimization, with RMSE decreasing from 0.4898 to 0.2009. Chlorophyll a, water temperature, and ammonia nitrogen are the key factors driving DO variations. SHAP LOWESS interpretable analysis identified the transition intervals for the empirical effects of temperature and salinity (water temperature 95% CI: [21.69, 22.26] °C; salinity 95% CI: [32.79, 33.16] psu), indicating a unique nonlinear response pattern for DO in the semi-enclosed Victoria Harbor. Independent results from multiple sites demonstrate that the MFSM framework improves the estimation performance of feature-redundancy-sensitive models (Random Forest) across multiple subregions of Victoria Harbor (RMSE reduction ranging from 31.1% to 59%). The adaptive modeling framework developed in this study addresses the issue of blind configuration in traditional DO estimation and provides reliable technical support for accurate water quality estimation, ecological association analysis, and intelligent estimation and control in coastal semi-enclosed bays.