Prediction heavy metals accumulation risk in rice using machine learning and mapping pollution risk

Rapid and accurate prediction of metal bioaccumulation in crops are important for assessing metal environ-mental risks. We aimed to incorporate machine learning modeling methods to predict heavy metal contents in rice crops and identify influencing factors. We conducted a field study in Jiangsu province, China, collecting 2123 pairs of soil-rice samples in a uniform measurement and using 10 machine learning algorithms to predict the uptake of Cd, Hg, As, and Pb in rice grain. The Extremely Randomized Tree model exhibited the best per-formance for rice-Cd and rice-Hg (Cd: R2 = 0.824; Hg: R2 = 0.626), while the Random Forest model performed best for As and Pb (As: R2 = 0.389; Pb: R2 = 0.325). The feature importance analysis showed that soil-Cd and pH had the highest impact on rice-Cd risk, which is in line with previous studies; while temperature and soil organic carbon were more important to rice-Hg than soil-Hg. Then, based on another set of 1867 uniformly distributed paddy soil samples in Jiangsu province, the Cd and Hg risks of soil and rice were visualized using the established models. Mapping result revealed an inconsistent pattern of hotspot distribution between soil-Hg and rice-Hg, i.e., a higher rice-Hg risk in the northern area, while higher soil-Hg in south. Our findings highlight the importance of temperature on Hg bioaccumulation risk to crops, which has often been overlooked in previous risk assessment processes.

Automated summary

This study used machine learning models to predict heavy metal contents in rice crops and identified influencing factors. The Extremely Randomized Tree model showed the best performance for predicting Cd and Hg, while the Random Forest model was the best for As and Pb. The feature importance analysis revealed that soil-Cd and pH had the highest impact on rice-Cd risk, and temperature was important for rice-Hg bioaccumulation risk.