Machine Learning Enables Low-Cost Determination of Soil Heavy Metal Concentrations

Soil heavy metals and metalloids (heavy metal[loid]s) are detrimental to human and ecosystem health. However, their concentrations are not regularly tested in many regions because of the dangerous and time-consuming pseudo extraction methods required for chemical analysis. In comparison, the identification of exchangeable heavy metal(loid) concentrations is advantageous regarding risk and costs. This study demonstrates that the total heavy metal(loid) concentration can be accurately predicted using the exchangeable concentration, elemental descriptors, and soil properties. A systematic modeling process was used to develop the ExtraTrees model, which was built on a global data set containing 9557 data points. ExtraTrees achieved a testing R2 of 0.9 for predicting the total concentrations. Multiple model interpretation tools revealed that the exchangeable concentration and magnetic moment of the element were the key drivers for estimating total concentration. The relative soil arsenic (As) threshold map suggested that stricter thresholds should be adopted in southeastern and north-central China considering their potential susceptibility to soil As contamination based on safe grain production. This study presents a rapid and efficient method for evaluating heavy metal(loid) concentrations that eliminates the need for dedicated instruments. These findings provide new insights into identifying heavy metal(loid) contamination while offering valuable guidance for developing evidence-based contamination policies, formulating effective remediation strategies, and implementing cleaner crop production practices.