In this study, the performances of different models were compared in spatial prediction of soil organic matter (SOM) in a karst region of south subtropical China. SOM content data were collected from the soil testing and formula fertilization in Longan County, and divided into karst areas and non-karst areas affected by karst areas. SPSS correlation analysis and R language random forest were used to screen environmental variables, eight kinds of digital soil mapping models were constructed and used for SOM prediction, including ordinary kriging (OK), inverse distance weighting (IDW), linear regression (LR), regression kriging (RK), geographical weighted regression (GWR), geographical weighted regression kriging (GWRK), random forest and random forest kriging (RFK). The results showed that:1) After the partition of the karst and no-karst areas, the prediction accuracy of karst region was obviously higher than that of non-karst region; 2) By ranking the importance of environmental variables, it was found that soil type, land use type, organic fertilizer dosage, drainage capacity, long-term precipitation, elevation, and irrigation capacity were the main environmental variable factors affecting the spatial distribution of SOM in karst areas; 3) The overall prediction accuracy from high to low was ranked as RFK>RF>RK>GWRK>GWR>IDW>LR>OK for the karst areas, with the R²=0.572 for RFK; while as RK>RFK>RF>LR>OK>GWRK>GWR>IDW of the non-karst areas, with the R²=0.439 for RK; 4) The prediction accuracies of LR, GWR and RF were improved to some extent after ordinary kriging interpolation of the residuals, indicating that SOM had a certain spatial correlation in space. The above results can provide technical reference for the high precision mapping of SOM in karst area of south subtropical regions.