Predicting Cd accumulation in crops and identifying nonlinear effects of multiple environmental factors based on...

Area of Science:

• Environmental Science
• Agricultural Science
• Geochemistry

Background:

• Traditional prediction of cadmium (Cd) content in crops (Cd_g) uses linear regression based on soil Cd (Cd_s) and pH, neglecting complex interactions.
• Existing models fail to account for nonlinear relationships and the influence of broader environmental factors on Cd accumulation.

Purpose of the Study:

• To develop and validate advanced machine learning models for accurate prediction of Cd_g in rice and wheat.
• To construct a comprehensive index system incorporating soil, geological, climatic, and anthropogenic factors.
• To compare the predictive performance of machine learning models against traditional linear regression.

Main Methods:

• Developed a comprehensive environmental factor index system.
• Employed machine learning models: tree-based ensemble (XGboost, random forest), support vector regression, and artificial neural network.
• Validated models using a test dataset and compared R² values with traditional linear regression models.

Main Results:

• Machine learning models significantly outperformed traditional linear regression in predicting Cd_g.
• XGboost and random forest achieved the highest accuracies (R² = 0.349 for rice, R² = 0.546 for wheat).
• Soil properties (Cd_s, pH, clay) were key drivers, but geological and climate factors significantly impacted wheat Cd_g due to regional differences.

Conclusions:

• Machine learning offers a superior framework for predicting crop Cd content by integrating diverse environmental data.
• The non-linear relationship between Cd_s and Cd_g highlights the complexity of cadmium uptake.
• This approach provides a novel, high-precision method for optimizing soil-plant transfer models and managing crop cadmium contamination.