Interpretable machine learning models to predict cadmium in wheat for safe production and soil management

Area of Science:

• Agricultural Science
• Environmental Science
• Data Science

Background:

• Cadmium (Cd) contamination in wheat grain poses risks to food safety and agricultural sustainability.
• Accurate prediction of Cd accumulation in wheat is crucial for risk assessment and management.

Purpose of the Study:

• To develop and validate a machine learning (ML) model for predicting wheat grain cadmium concentrations based on soil properties.
• To identify key soil factors influencing Cd uptake in wheat and establish soil Cd safety thresholds.

Main Methods:

• Utilized a dataset of 1,339 soil-wheat grain pairs to train and evaluate nine ML algorithms.
• Employed eXtreme Gradient Boosting (XGBoost) for its superior predictive performance and Shapley Additive Explanations (SHAP) for feature importance analysis.
• Validated the model with nation-scale data and developed an online application for practical use.

Main Results:

• The XGBoost model achieved high predictive accuracy (R² = 0.90), significantly outperforming multiple linear regression (R² = 0.69).
• Soil total Cd and pH were identified as the primary drivers of wheat grain Cd accumulation, with soil Cd having a positive effect and pH a negative one.
• High-risk areas for Cd accumulation were pinpointed in China, and soil Cd safety thresholds were proposed based on soil pH.

Conclusions:

• Interpretable ML models, particularly XGBoost, offer a robust tool for predicting wheat grain Cd and managing soil contamination.
• The developed model and online application provide actionable insights for ensuring food safety and promoting sustainable wheat production.
• Establishing soil Cd safety thresholds is essential for compliance with regulatory limits and mitigating risks associated with Cd accumulation.