The soil samples were collected with gridding method based on a dynamic monitoring plot (200 m × 40 m) in manmade forest of karst peak-cluster-depression region. By the methods of classical statistics and geostatistics, spatial heterogeneity of main mineral composition (SiO2，Fe2O3，CaO，MgO，Al2O3，MnO) in surface soil (020 cm) was studied. The differences of the six mineral compositions were large with SiO2 having the highest contents followed by Al2O3, Fe2O3, MgO, CaO, and MnO and the variance coefficients ranged from 6.90% to 28.89%. Spherical model fitted best for Fe2O3, CaO, and MgO, Gaussian model best for SiO2 and MnO, and Exponential model best for Al2O3. MgO and CaO had medium spatial autocorrelation, while the others were characterized by strong spatial autocorrelation. The Kriging contour map of CaO was similar with that of MgO. The Kriging contour maps of SiO2 and MnO were similar. Fe2O3 and Al2O3 presented spatial distribution of fragmented patches.