Applying machine learning methods to resolve the cadmium (Cd) uptake characteristics of regional soil-wheat systems can contribute to the accuracy and rationality of risk decisions. Based on a regional survey, we constructed a Freundlich-type transfer equation, random forest (RF) model, and neural network (BPNN) model to predict wheat Cd enrichment factor (BCF-Cd); verified the prediction accuracy; and assessed the uncertainty of different models. The results showed that both RF (R2=0.583) and BPNN (R2=0.490) were better than the Freundlich transfer equation (R2=0.410). The RF and BPNN were further trained repeatedly, and the results showed that the mean absolute error (MAE) and root mean square error (RMSE) of RF and BPNN were close to each other. Additionally, the accuracy and stability of RF (R2=0.527-0.601) was higher than that of BPNN (R2=0.432-0.661). Feature importance analysis showed that multiple factors led to the heterogeneity of wheat BCF-Cd, in which soil phosphorus (P) and zinc (Zn) were the key variables affecting the change in wheat BCF-Cd. Parameter optimization can further improve the accuracy, stability, and generalization ability of the model.
Keywords: neural network; random forest; regression equation; uptake factors; wheat.