PRISM: Turning prediction uncertainty into cost-effective management decisions for cadmium-contaminated rice

Ensuring rice safety in cadmium (Cd) contaminated regions is critically challenged by the disconnect between predictive modeling and risk management. Conventional approaches prioritize point estimation accuracy but neglect predictive uncertainty and asymmetric decision costs, leading to costly misclassification errors where safe soils produce unsafe rice or safe fields undergo unnecessary remediation. Current static threshold-based standards further fail to capture the complex biogeochemical modulation of Cd bioavailability. To bridge this gap, we propose PRISM (Probabilistic Risk Integrated Spatial Management), a framework integrating Bayesian inference, bagging ensembles, and quantile regression to replace deterministic estimates with probabilistic outputs. Applied to a typical rice-producing region in China, SHAP analysis reveals a source-valve mechanism: soil properties (CEC, organic carbon density) act as Cd sources, while moisture-regulating topographic and climatic factors serve as valves controlling bioavailability. Crucially, under an asymmetric loss matrix weighing health risks against economic costs, PRISM reduces total expected management costs by 65.2% and 18.6% compared to conventional soil and rice standard strategies, respectively. Beyond cost savings, this study demonstrates that high-uncertainty zones should trigger low-cost investigations rather than being treated as noise. We argue that for food-safety-oriented environmental management, quantifying model confidence is fundamentally more decision-relevant than pursuing marginal gains in point prediction accuracy, offering a generalizable paradigm for resource allocation under uncertainty.