Abstract:In order to improve the accuracy of spatial interpolation of soil nutrients in farmland and accurately grasp the spatial distribution characteristics of soil nutrients, variable screening were performed by using Pearson correlation coefficient, variance inflation factor and extreme gradient boosting algorithms. Then, decision tree, random forest, radial basis function and long short-term memory were used with ordinary Kriging to interpolation the content of soil nutrients in the farmland. The results showed that the soil organic matter, total nitrogen, available phosphorus, and available potassium contents in the study area ranged from 0.226 g/kg to 32.275 g/kg, 0.117 g/kg to 1.272 g/kg, 3.159 mg/kg to 53.884 mg/kg, and 81.510 mg/kg to 488.422 mg/kg, respectively, with moderate variability. PCC, VIF and XGBoost variable screening all showed that soil organic matter, total nitrogen, available phosphorus and available potassium had some correlation among them and can be used as environmental variables for the spatial interpolation of target attributes. XGBoost method can more effectively screen out the environmental variables that were important to the spatial interpolation results, and the accuracy of the model built after screening variables based on this method was significantly better than the accuracy of the model built after screening variables by PCC and VIF. Moreover, the accuracy of the machine learning model with the synergistic environmental variables was generally better than the accuracy of the OK model without environmental variables, and the accuracy of the spatial interpolation model for the same soil nutrient content showed the following order: RF>LSTM>RBF>DT>OK. Using the RF model to invert soil nutrients in the study area, it was found that the soil organic matter and total nitrogen higher content was mainly concentrated in the southern and eastern regions of the study area, the available phosphorus and available potassium lower content in the southeastern and north-central regions. In summary, the XGBoost variable screening method combined with RF model can better realize the spatial interpolation of soil nutrients, and can be used as an effective method for the spatial interpolation of soil nutrients.