Soil microbial biomass (SMB) is considered to be an important indicator of soil fertility and biological quality, and it presents strong spatial heterogeneity in relation to soil properties (e.g. soil texture, organic carbon, total nitrogen) and topography at various spatial and temporal scales. In the present study, geostatistics were used to analyze the spatial variability of soil microbial biomass carbon (MBC), nitrogen (MBN) and phosphorus (MBP) in a hilly red soil landscape (446 ha) in subtropical China. Five hundred and twenty-three soil samples at a soil depth of 0–20 cm were randomly collected from paddy fields, dry lands, orchards and wood lands in the study region. Significant negative correlations (r = −0.69 to −0.54) of MBC, MBN and MBP with elevation were observed. The Stein’s Matérn, Spherical and Gaussian models with effective ranges of 157, 252 and 213 m were best-fitted to the sample semivariograms of original MBC, MBN and MBP variables, respectively. All three SMB variables exhibited moderate spatial dependence. After detrending the elevation impact, the normal score transformed MBC and MBP still retained their moderate spatial autocorrelations with slightly decreased effective ranges, whereas the normal score transformed MBN demonstrated a very strong spatial dependence with a much shorter effective range of 70 m. Spatial distributions of the three SMB variables were estimated using both ordinary kriging (OK) and regression kriging (RK) with elevation as the predictor. The kriging predictions showed that soil MBC, MBN and MBP had overlapping spatial patterns, and furthermore the RK interpolations showed more details in space than the OK interpolations, with improved prediction accuracy. The spatial distribution of soil MBN demonstrated more hotspots than the other two SMB variables, implying that MBN might be more sensitive to environmental disturbances (such as fertilization, tillage and crop rotations).