Climate-responsive assessment of surface water quality in Songhua River using ensemble learning and multivariate analysis

Abstract

Accurately identifying and regulating water quality drivers is vital for sustainable management, but complex climatic, hydrological, and pollution interactions pose significant challenges. Herein, we propose an integrated prediction framework combining an improved Water Quality Index (WQI), multivariate analysis, and a self-adaptive ElasticForest model. The proposed framework is applied to the Songhua River Basin in China to evaluate and predict surface water quality under changing climatic conditions. The improved WQI method assesses water quality, followed by cluster analysis identifying two representative hydrological periods: ice period (IP) and wet period (WP). Principal component analysis and factor analysis extract major pollution sources, while Random Forest selects key water quality indicators for each stage. The proposed ElasticForest model combines the nonlinear learning capacity of RandomForest with the regularization strength of ElasticNet. This integration allows for efficient variable selection and robust prediction under conditions of high dimensionality, limited samples, and multicollinearity. By including water temperature and rainfall, the model enhances its sensitivity to climate-driven changes in water quality and achieves high prediction accuracy (R² = 0.978 for IP; R² = 0.989 for WP), outperforming both traditional and ensemble models in stability and generalizability. Ultimately, the relationship between sediment heavy metals and water parameters is analyzed. High-risk metals like Hg, Cd, and Ni show positive correlation with TN, DO, and BOD, while Cu and Zn are negatively correlated with nutrients. These findings suggest sediments shape pollution dynamics via biogeochemical processes. The study underscores climate-water-sediment interactions and offers a framework for climate-sensitive, interpretable water quality prediction.