Machine learning for monitoring and assessment of potentially toxic elements in soils: a synthesis of spatial validation, explainability, and uncertainty.

Potentially toxic elements (PTEs) in soils pose persistent risks to ecosystems, groundwater, and food systems, creating a need for reliable spatial assessment tools. Machine learning (ML) is increasingly used to map PTE concentrations from environmental covariates, but many studies still rely on spatially naive validation, limited interpretation, and incomplete uncertainty reporting. This review synthesizes recent advances (2020-2025) in ML-based PTE mapping with emphasis on four requirements for monitoring-grade assessment: spatially honest validation, explainable artificial intelligence, uncertainty quantification, and translation of predictions into decision-ready risk products. We show that random cross-validation often overestimates predictive performance when spatial dependence is ignored and summarize emerging block-based and dissimilarity-aware evaluation strategies. We also assess the use of Shapley Additive exPlanations (SHAP) and related tools for identifying geogenic and anthropogenic drivers, and we review ensemble, quantile, and probabilistic methods for producing exceedance-probability maps aligned with regulatory thresholds. Building on this synthesis, we propose a reporting checklist and a risk-uncertainty decision matrix to improve reproducibility, transparency, and policy relevance. Overall, predictive accuracy alone is insufficient for environmental decision-making; robust validation, interpretable modeling, and uncertainty-aware risk translation are essential for scientifically defensible soil-contamination management.