Machine learning and remote sensing for soil heavy metal(loid) risk assessment: A systematic review of multi-source data fusion, benchmarking, and explainable AI.

Soil heavy metal(loid) contamination threatens food security and public health, but conventional monitoring alone cannot provide the spatial coverage required for effective risk assessment and management. This PRISMA-compliant review synthesizes 245 eligible reports published between 2015 and 2025 on machine learning and remote sensing for soil contamination prediction and risk assessment. Across 536 eligible conventional R² estimates from 200 reports, the field-wide mean was 0.722 at the estimate level and 0.749 at the report level. Algorithm-family differences were descriptive and confounded by target element, sample size, input data, study context, validation design, and reporting quality. No report used spatial block cross-validation; 99 (40.4%) used random train-test splitting, 38 (15.5%) used random k-fold cross-validation, and 72 (29.4%) inadequately documented partition construction. Restricting analyses to documented protocols reduced the estimate-level mean R² to 0.707. Interpretability remained limited: 82 reports (33.5%) used no method, 46 (18.8%) provided insufficient documentation, and 19 (7.8%) used SHAP, which should be interpreted as a model attribution method rather than a causal inference method. Only 43 reports (17.6%) included health or ecological risk assessment, and 11 (4.5%) explicitly applied EPA HQ/HI frameworks. A six-item Minimum Reporting Standard is proposed to strengthen validation, transparency, interpretation, and risk integration.