Taking Yingtan City, Jiangxi Province as the study area, the topsoil (0-20cm) bulk densities in 131 sampling sites under different land uses and soil types were investigated, 66 environmental factors such as topography, remote sensing, climate and so on were extracted to form the original feature set by combining the multi-source environmental big data. And a method of multiple am plification and screening of the feature variables was created, that was, for the original feature set and random forest (RF) model, PCA component extraction - component augmentation - recursive feature elimination with cross validation (RFECV) - polynomial feature expansion (PFE) - recursive feature elimination with cross validation (RFECV) re-screening being carried out in order. In the end, the optimal set of features for the combination of 3 variables was obtained. The spatial prediction accuracy of RF for soil bulk density based on the optimal feature set reached R² of 0.469, which was 34% higher than that of the original feature set (R²=0.315), and the feature dimensionality was reduced by 95%, which significantly improved the spatial prediction effect and efficiency.