Abstract: The strong spatial heterogeneity of soil environmental variables causes difficulties in improving spatial interpolation accuracy. It is difficult to obtain a high interpolation accuracy by leveraging spatial correlation and spatial heterogeneity. Machine learning methods can fuse the information of multi-dimensional auxiliary variables to improve the interpolation accuracy of soil attributes, but they cannot effectively utilize the spatial position relationship information to further improve the interpolation accuracy. Based on the random forest spatial prediction framework, this study combined the spatial semi-variogram with the random forest algorithm and proposed a spatial random forest interpolation method with a semi-variogram. Taking soil heavy metal data from the Xiangtan County of Hunan Province as an example, the proposed method was used to implement spatial interpolation of soil Cr. The interpolation accuracy was compared with the random forest method, distance-based random forest spatial prediction method, ordinary Kriging method, and regression Kriging method. The results showed that the accuracy was improved by more than 10% compared with the traditional Kriging method. Compared with the new machine learning spatial interpolation method, the accuracy was improved by more than 5%. Furthermore, the mapping of the proposed results had a more reasonable spatial distribution and detailed information. Thus, we concluded that the proposed method could effectively combine auxiliary variable information and spatial location information and improve the interpolation accuracy of soil environmental variables.