New soil models improve safety of wheat amid cadmium contamination

A new study offers a new solution to tackle cadmium contamination in wheat, one of China’s most vital staple crops. Scientists have developed a suite of predictive models that estimate cadmium levels in wheat grain using soil characteristics. Among these, a machine learning model dramatically outperformed traditional methods, enabling faster and more precise risk assessments. By analyzing 311 soil-wheat samples collected from across China, the researchers identified soil pH, cadmium content, and cation exchange capacity as key drivers of metal uptake. The study also introduced improved soil cadmium thresholds that better reflect food safety needs, offering a practical path toward safer and more sustainable wheat production.

Cadmium (Cd), a toxic heavy metal, poses a growing threat to food safety through its accumulation in crops. Wheat, in particular, tends to absorb more cadmium than rice due to its higher internal transport efficiency. In China’s rice–wheat rotation systems, wheat grains often exceed cadmium safety limits, despite adherence to national soil quality standards. Recently revised regulations have introduced stricter soil cadmium limits, but they may result in unnecessary costs and over-regulation. Earlier predictive models have fallen short, failing to capture the complexity of real soil conditions. Due to these challenges, there is a pressing need for more accurate, field-validated models to ensure wheat safety without overburdening producers.

On May 14, 2025, researchers from Nanjing University and Columbia University published a study (DOI: 10.1016/j.eehl.2025.100154) in Eco-Environment & Health that unveils new models to predict cadmium accumulation in wheat grain. Using data from 311 paired soil and wheat samples across China, the team compared multiple regression, machine learning, and geochemical approaches. Their aim was to pinpoint the most effective model and generate precise soil cadmium thresholds tied to national food safety standards—offering a more informed framework for protecting wheat from contamination.

The team identified soil total cadmium, pH, and cation exchange capacity (CEC) as the most influential factors in cadmium uptake by wheat. Based on these variables, they built predictive models, including one that used CaCl₂-extractable cadmium to represent the bioavailable fraction most relevant to plant absorption. A geochemical model—the Multi-Surface Speciation Model (MSM)—was also tested for its ability to simulate cadmium behavior under varying soil conditions. While both methods performed well, the standout was the Extremely Randomized Trees (ERT) machine learning model. It achieved a root mean square error (RMSE) of 0.221 and mean absolute error (MAE) of 0.165, outperforming other models in accuracy and adaptability. Crucially, the researchers used these models to back-calculate soil cadmium thresholds based on China’s food safety limit of 0.1 mg/kg for wheat grain. These newly derived thresholds—adjusted for different soil pH levels—proved more effective in predicting grain safety than current national standards, offering a refined and cost-efficient alternative to blanket soil remediation.

"Our goal was to create a practical tool that farmers and regulators can use to assess wheat safety directly from soil data," said Professor Xueyuan Gu, corresponding author of the study. "The machine learning models and new thresholds we developed are not just academic exercises—they can be integrated into field management systems and national monitoring programs." She emphasized the importance of combining scientific rigor with practical usability, noting that broader data collection across regions could further improve the model's reliability and generalizability.

This research has significant implications for agricultural safety and policy development. With these models, cadmium risks can be assessed rapidly and accurately using standard soil tests, empowering farmers and local authorities to make informed decisions about land use. The refined thresholds provide a science-based, economically feasible alternative to rigid remediation policies, helping prevent both under- and over-regulation. Moreover, the successful integration of machine learning marks a broader shift toward data-driven agriculture. As soil databases expand, these predictive tools could evolve into real-time advisory systems—enhancing sustainable land management while protecting public health through safer food production.

###

References

DOI

10.1016/j.eehl.2025.100154

Original Source URL

https://doi.org/10.1016/j.eehl.2025.100154

Funding information

This work was financially supported by the Natural Science Foundation of China (42177188, 21876080).

About Eco-Environment & Health

Eco-Environment & Health (EEH) is an international and multidisciplinary peer-reviewed journal designed for publications on the frontiers of the ecology, environment and health as well as their related disciplines. EEH focuses on the concept of "One Health" to promote green and sustainable development, dealing with the interactions among ecology, environment and health, and the underlying mechanisms and interventions. Our mission is to be one of the most important flagship journals in the field of environmental health.