Soil fuzzy membership values of the sampled soils resulted from fuzzy c-means algorithm, which would be applied in soil predictive mapping, is a kind of compositional data. Owing to the structural characteristics of compositional data, they could not be directly used in prediction of the fuzzy memberships of unknown sites over space by kriging interpolation. To achieve spatial soil prediction, therefore, the membership values of the sampled soils must be transformed before interpolation. In this study, transform of compositional data by several ways were attempted, and influence of different transform approaches on output and precision of prediction compared and analyzed. The results indicated that, membership values of all the spatial predicted sites didn't sum to 1 on condition that known membership values of the sampled soils were simply transformed through logarithm. Obviously, the above predictive result was theoretically unauthentic. Contrarily, membership values of all the spatial predicted sites summed to 1 when the membership values of the known soils were transformed by asymmetry Logratio and symmetry Logratio approaches, demonstrating that two approaches were theoretically accepted. Comparatively, symmetry Logratio transform could lead to a better spatial distribution pattern and higher precision.