Rainfall-dependent influence of water parameter interactions and land use on lake water quality: A hybrid ensemble approach and management implications

Understanding how water parameter interactions and land use affect lake water quality is crucial for water management and ecological sustainability. However, rainfall patterns alter these impact characteristics, an issue that has rarely been addressed systematically. This study established a methodological framework combining Bayesian Network-Extreme Gradient Boosting (BN-XGB) ensemble models to describe rainfall-dependent influences of natural and anthropogenic factors on water quality, with model structures identified through Structural Equation Modeling (SEM) and interpreted using Shapley Additive Explanations (SHAP). The methodology was demonstrated in Hongze Lake, China, using daily water quality data from six monitoring stations. Results showed that the BN-XGB model outperformed standalone models, with correlation coefficients of 0.79–0.91 and Nash-Sutcliffe efficiency coefficients of 0.61–0.82. Complex temporal patterns were captured showing oxygen (DO) responding to short-term rainfall (within 3 days), nutrients to medium-term patterns (7–15 days), and algal indicators to antecedent dry days. Positive effects of DO on Ammonia Nitrogen (AN) weakened with increased rainfall, while opposite effects existed for Total Nitrogen (TN) and Total Phosphorus (TP). DO dominated algal growth (90 %) during drought periods, whereas prolonged rainfall shifted the control to accumulated nutrients. Artificial surfaces, water bodies and farmland acted as pollution sinks during low rainfall but became sources as rainfall increased, contributing over 40 % to nutrients and algal growth. However, these areas transformed into pollution sinks after heavy rainfall or storms. While DO and AN were insensitive to land use changes, TP and TN severely deteriorated in agriculture/urban-dominated scenarios (>70 % and 45 % of records worse than Class V), and eutrophication risks increased when farmland exceeded 60 %. Water quality improved under ecological protection scenarios and reached optimal conditions in balanced scenarios. This methodological framework provides replicable theoretical support for scientific lake management and is applicable to polluted lakes globally.