Objective In order to accurately predict the spatial distribution characteristics of heavy metal content in agricultural soils, the applicability and uncertainty of different prediction methods were explored, so that an optimal model applicable to the spatial prediction of heavy metals in agricultural soils could be established.

Method Taking 0 ~ 20 cm of soil around a metal smelting plant in the Hetao area of western Inner Mongolia as the research object, the prediction methods, such as ordinary kriging (OK), random forest (RF) and ordinary kriging + random forest (RFRK) were adopted, and the environmental variables, such as soil physicochemical properties, topography, climate, anthropogenic activities and so on were selected as modelling covariates to construct a suitable model for the spatial prediction of heavy metals in agricultural soils, to reveal the characteristics of the spatial distribution of the heavy metal contents of As, Pb, Cr and Cd, and to explore the differences in the spatial prediction accuracy of the different modelling methods.

Result ① The average values of soil Cr, Pb and Cd are 3.5, 0.66 and 0.23 times higher than the background values of the soils in the Loop, and there is a certain degree of enrichment of heavy metals in the agricultural soils in the area, but it does not reach the screening value of soil risk in the agricultural land, and the content of heavy metals in soils in the northern and central parts of the study area, as well as in the area of the smelter, is higher than that of the other parts of the area. ② The prediction accuracies of soil As, Pb, Cr and Cd contents of OK, RF and RFRK models were in the order of RFRK > RF > OK, i.e. the RFRK model had the smallest mean absolute error (MAE) and root mean square error (RMSE) in the predicted values, and the cross-validated R2 values were all over 0.8, which showed the strongest correlation between the measured and predicted values.

Conclusion The Ordinary Kriging + Random Forest model could be used as an effective method for spatial prediction of heavy metal content in soil, and is priority for application in the investigation, assessment and prevention of soil pollution in regional farmland.