【Objective】 This study aimed to construct a high precision prediction method for soil bulk density to accurately complete the regional soil attribute database.【Method】 Based on the data of 2, 883 typical cropland samples in the Sichuan Basin (including Sichuan Province and Chongqing Municipality) obtained during the second national soil census, this study used correlation analysis, variance analysis, and regression analysis to reveal the statistical characteristics and main controlling factors of the cropland topsoil bulk density in the Sichuan Basin. The traditional pedotransfer functions (PTFs), multiple linear regression (MLR) models, radial basis function neural network (RBFNN) model, and random forest (RF) models were used to establish a soil bulk density prediction model through three modeling methods: whole region, by river basin and by soil type, to fill the missing value of soil bulk density.【Result】 The results show that the cropland topsoil bulk density in the study area ranged from 0.60 to 1.71 g·cm^-3, with a mean value of 1.29 g·cm^-3. Soil organic matter, soil subgroup, and rainfall in summer were the most important factors influencing bulk density. The RBFNN model constructed by the river basin can better capture the nonlinear relationship between soil bulk density and the influencing factors and the spatial non-stationarity of this relationship. The coefficient of determination (R²) and root mean square error (RMSE) of the 432 independent validation samples were 0.519 and 0.095 g·cm^-3, respectively, which were significantly better than those of other methods.【Conclusion】 Therefore, the RBFNN prediction model constructed in sub-basin is helpful to improve the imputation accuracy of the missing values of topsoil bulk density in the Sichuan Basin, and also provides a method reference for the imputation of missing values of soil properties in other regions.