A machine learning and counterfactual reasoning-integrated decision framework for cadmium remediation: Predictive...

Environmental remediation strategies for cadmium (Cd)-contaminated rice paddies often face challenges due to reliance on time-consuming field trials and limited pre-assessment of intervention efficacy. Here, we propose a machine learning and causal inference-integrated framework to enable proactive decision-making, using iron plaque-mediated Cd immobilization as a model system. By analyzing 76 paired soil-rice samples, extreme gradient boosting (XGBoost) and SHapley Additive exPlanations (SHAP) identified six critical drivers of grain Cd accumulation from 31 physicochemical and microbial indicators. Structural equation modeling revealed iron plaque reduced root Cd accumulation by 38.6 %, primarily regulated by available Fe (41.9 %) and pH (37.7 %). Counterfactual scenario simulations further quantified threshold combinations of Fe availability (300-400 mg·kg⁻¹) and pH (>5.5) to maintain grain Cd below 0.2 mg·kg⁻¹. Economic analysis in five towns of Xiangtan country demonstrated region-specific optimization: FeSO₄ amendment generated 1.37 × 10⁶ CNY·km^-2 net profit in soils with high native Fe content, while pH adjustment via Na₂SiO₃ achieved 2.68 × 10⁶ CNY·km^-2 in acidic regions. This framework bridges predictive analytics with cost-benefit evaluation, providing a paradigm shift from post hoc to preemptive remediation planning.