Spatial autocorrelation is the first law of geography. GISbased geostatistical analysis provides a scientific method to study the phenomenon of spatial autocorrelation and dependence. Geostatistical analysis of the major forest inventory factors in case study area of Zijinshan National Forest Park in 2002 showed that, at the spatial scale from 100 m to 1 000 m, there existed a common spatial correlation for the 10 investigated factors. With the increase of sample spacing, spatial autocorrelation coefficients of Moran I tended to decrease. Structural variance caused by space difference played a primary role in system variance, in which the proportions of structural variance of site factors(elevation, slope, aspect) to system variance were more than 90 % and other factors over 60 %. Spatial autocorrelation ranges of the three site factors were just 1 950 m, while the ranges of forest measuring factors of age, DBH, tree canopy density were bigger than 3 000 m. Interpolation accuracy analysis of stock volume per unit area of different theoretical models of semi variance showed that, among the five semivariance models of the linear, Gaussian, exponential, circular and sphere, the determination coefficient of exponential model was the highest(0.918), the correlation coefficient was the highest(0954), the residual standard deviation(21.438) and the average relative error(20.591 %) were the lowest. Evaluated from two aspects of the fitting accuracy and prediction precision, the exponential model outperformed all others, which was closely related with the patched forest landscape structure in the study area.