Artificial irrigation modulates heavy metal dynamics and risks in lowland agricultural ponds

Heavy metals (HMs) in aquatic ecosystems threaten environmental and public health, highlighting the urgent need for high-resolution, dynamic risk assessments. However, conventional methods are constrained by sparse monitoring data of HM concentrations and limited capacity to capture the cause-effect relationship between HM dynamics and environmental conditions. To address this gap, we developed a machine learning framework that integrates Shapley additive explanation (SHAP) analysis to describe daily dynamic of HM concentrations and risks, and to quantify environmental effects by incorporating dynamic coefficients. The framework was applied to a lowland agricultural pond during 2016-2019, and achieved robust performance (R²>0.69 during both training and testing periods) using random forest algorithm. Both simulated and observed data revealed an increasing trend during the study period, with seasonal peaks of arsenic (As), cadmium (Cd) and lead (Pb) in summer or autumn. Our analysis revealed diverse thresholds and interaction patterns of environmental factors governing HM enrichment, ranging from unidirectional (positive/negative) to bidirectional (U-shaped/inverted U-shaped) relationships, with artificial irrigation as the key driver. Daily risk assessments showed substantial temporal variability, with risk levels of low, moderate, considerable, and high accounting for 19%, 50%, 29%, and 2% of the study period, respectively. Notably, overall HM risk was driven predominantly by human carcinogenic risks, which remained present even at levels compliant with water quality guidelines. This study demonstrated the value of the proposed framework in enabling fine-scale risk assessment and improving mechanistic understanding, thereby offering practical benefits for heavy metal (HM) risk control in water management.