Abstract: In order to establish regional models for describing soilrice grain cadmium (Cd) coupling relationship, we collected 369 groups of data from literature to construct models, which used generalized linear model (GLM), gradient boosting machine (GBM), random forest (RF) and Cubist methods, with soil pH and total Cd content (Soil_Cd) as independent variables and Cd content in rice grains (Grain_Cd) as the dependent variable. The robustness of those models in the Grain_Cd predictions was evaluated using the measured data of soil pH, Soil_Cd and Grain_Cd. Results showed that GLM, GBM, RF and Cubist models showed similar performance, with all of their coefficients of determination (R²) being around 0.5. The measured Grain_Cd values were best matched to the prediction of the RF model (R²=0.534). Therefore, the RF model, which is based on machine learning, was capable to reasonably predict Cd content in rice grains at the regional scale.

Key words: soil, rice, Cd, coupling relationship model, machine learning.