Accurately predicting the spatial distribution of soil organic matter (SOM) content is of great significance for improving soil quality and improving the level of regional soil management. In order to explore the optimal model for predicting the SOM content of farmland in the Weibei Dryland of Shaanxi Province, the influence factors closely related to SOM content were selected as the modeling covariables, and a geographic detector, the ordinary kriging method (OK), geographic weighted regression model (GWR), partial least squares regression model (PLS), geographically weighted regression extended model (GWRPLS), and random forest model (RF) were used to predict the spatial distribution of SOM content in training samples. Additionally, the validation set samples were used to compare and analyze the prediction accuracy of the five methods. The results showed:① the main factors affecting the spatial variability of soil SOM were total nitrogen, fertilizer application, available potassium, available phosphorus, and altitude, and the interaction between any two factors was more explanatory for SOM than any single factor. ②ω(SOM) in farmland was between 2.25 and 30.23 g·kg-1, with an average value of 15.14 g·kg-1 and a coefficient of variation of 30.00. Although there were local differences in the prediction results of SOM by the five methods, the overall spatial distribution trend was basically the same. In the study area, the content of organic matter was low in the north and northeast and high in the west and southeast. ③ From the perspective of the prediction accuracy of the five methods, the root mean square error (RMSE) and mean absolute error (MAE) of RF were the smallest, and the prediction deviation (RPD) of GWRPLS was the largest. Compared with the OK method, the correlation coefficients (r) of GWR, PLS, RF, and GWRPLS increased to 0.907, 0.836, 0.968, and 0.972, respectively. Comprehensive analysis results showed that the random forest model had the highest prediction accuracy.
Keywords: geographic detector; geographic weighted regression(GWR); random forest(RF); soil organic matter(SOM); spatial prediction.