Evaluation of machine learning models incorporating soil properties and speciation to determine soil bioavailable lead.

The contamination of farmlands with lead (Pb) has become a long-term threat to food safety and ecosystem health. Risk assessment based on the quantification of total Pb concentration is insufficient, as bioavailability is governed by complex non-linear interactions within the soil-crop system. This study aimed to predict bioavailable Pb from lead speciation and soil properties using machine learning. We compiled a global dataset consisting of paired soil-crop Pb data from the Web of Science database. Ten machine learning models were developed and validated on soil available Pb-crop Pb dataset. Hyperparameter optimization was performed using grid search with fivefold cross validation to select the optimal model. SHAP analysis was conducted to determine which features most strongly influenced model predictions and to quantify their individual effects. The optimized XGBoost model achieved the highest prediction with (R2 = 0.9), showing non-linearity in the system. Interpretable analysis revealed organic matter and pH as the most dominant controllers, and critically identified the acid-soluble fraction as the primary predictor of risk, despite the reducible fraction having a higher concentration. Pb accumulation in rice grains followed a linear relationship with soil available Pb (R2 = 0.419), while wheat exhibited a more complex, cubic uptake (R2 = 0.37). This study provides a robust, interpretable tool that goes beyond total Pb concentration to predict bioavailable Pb in soil. Our findings indicate that risk assessment measures should prioritize not only soil properties but also Pb fractions and crop-specific models to ensure the sustainable use of agricultural systems.