An integrated geochemical and explainable machine learning framework for mercury dynamics and risk assessment in the Musa Estuary.

Abstract

Mercury (Hg) contamination in estuarine environments is of increasing concern because of its persistence, transformation into methylmercury (MeHg), and potential ecological and human health impacts. This study investigated the distribution, environmental controls, and health risks of Hg and MeHg in 41 surface sediment samples and 25 water samples collected during a single sampling campaign from the Musa Estuary (southern Iran) by integrating geochemical assessment with explainable machine learning. Sediment total mercury (THg) concentrations ranged 10.97 to 8,367.38 μg/kg, while MeHg concentrations ranged from 0.405 to 9.058 μg/kg. Geochemical indices revealed severe anthropogenic mercury contamination at several sampling sites. Three machine learning algorithms, Random Forest (RF), Gradient Boosting (GB), and Extreme Gradient Boosting (XGBoost), were evaluated to predict mercury concentrations. GB achieved the highest predictive performance for sediment THg (R2 = 0.874) and MeHg (R2 = 0.870), whereas XGBoost performed best for methylation potential (R2 = 0.789) and THg in water (R2 = 0.815). SHapley Additive exPlanations (SHAP) identified organic matter as the dominant driver of THg accumulation in sediments, whereas pH primarily controlled MeHg formation and THg concentration in water. Health risk assessment indicated that Hg posed no unacceptable non-carcinogenic risk under the assessed exposure scenarios (HI < 1); however, sediment-associated Hg exposure was substantially higher than water exposure, with children showing the greatest estimated risk (sediment HI = 1.65 × 10-2). This study demonstrates that integrating geochemical assessment with explainable machine learning provides an effective framework for understanding mercury dynamics and supporting risk-based management of contaminated estuarine ecosystems.