Explainable and causal machine learning to investigate the spatiotemporal dynamics patterns of coastal water quality in Hong Kong

• Developed an explainable–causal ML framework for coastal water quality attribution • Model prediction importance is not necessarily equivalent to causal influence • Rising human pressure undermines the natural mitigation of algal levels • Agriculture-driven eutrophication aligns with colder, faster-flowing, saline waters Coastal marine water quality emerges from complex and dynamic feedbacks between natural processes and human activities, yet disentangling their respective influences remains a persistent challenge. This study proposes an interpretable, data-driven framework that integrates clustering, explainable machine learning (XML), and causal inference to investigate long-term coastal water quality dynamics. Using 36 years of monthly in-situ measurements of ten water quality parameters from 76 monitoring stations across ten water control zones (WCZs) in densely urbanized Hong Kong, we clustered into several representative water quality regimes with distinct spatiotemporal characteristics: low pollution, microbial, and eutrophic. XML-based SHAP analysis revealed regime-specific patterns: microbial pollution was best predicted by urban proximity, clean waters by shipping intensity, and nutrient enrichment by agricultural land use, with salinity and SiO 2 consistently ranking as dominant environmental drivers. Building on SHAP results, we applied CausalForestDML to estimate the marginal effects of human drivers while controlling for a comprehensive set of environmental confounders. While SHAP importance generally aligned with causal effects, notable discrepancies underscored the added value of causal modeling. Further, temporal causal analysis revealed attenuated urban and shipping influences on DO, while agricultural impacts on nutrient concentrations have intensified and become more spatially heterogeneous. Although proximity to natural land consistently mitigated E. coli , its buffering capacity on nutrient pollution appears to be weakening under expanding agricultural pressure. The proposed framework demonstrates how combining predictive and causal analytics can integratively reveal mechanistic insights into water quality evolution and regulation, offering transferable tools for sustainability-oriented coastal governance.