Provided is a pedo-transfer function (PTF)-based method for predicting a target soil property. Based on the collection of a multi-source soil dataset and environmental variables, a dataset containing all measured information is divided. Second-level regions are obtained by zoning according to the spatial variation in soil properties. An optimal independent variable set of PTFs in different regions is obtained by screening. Then, linear fitting and nonlinear fitting of the PTFs are performed for different zones separately. By comparing the accuracy of different functions between different zones, optimal PTFs oriented toward sampling sites are selected, so as to build a database including soil sampling sites. Further, regional independent variable layers are constructed by means of machine learning, to establish region-oriented PTFs; and a spatial distribution map of the target soil property and content for a target region is produced.