【Objective】It is of great practical significance to investigate, manage and exploit soil resources based on knowledge mining and updating of historical soil maps, while screening of representative training samples out of the historical soil maps is a key step to accomplish the task. 【Method】Jingde County of Xuancheng City in Anhui Province was cited as the study area. In this paper, a new method for screening of training samples was developed, consisting of determining quantity of samples and specifying sample locations. The area segmented linear-scaling method, which builds a linear correspondence between the area and the number of samples in each area segment, was applied to determination of quantity of sample in each historical soil map polygon, and then after making geographical neighborhood analysis of elevation and slope, the two important environmental covariates, the stable cells in spatial variation of the environmental covariates were defined as sample locations. Geographical neighborhood analysis index indicates the degree of spatial variation of the environmental covariates in the neighborhood.【Result】Results show that the use of the area segmented linear-scaling method to determine quantity solved the problem that had never been pondered in past researches of how to distribute samples among units that were of the same type, but consisted of a number of polygons. When this method was used to define sample locations in polygons located in topographically flat areas, determination of spatial distribution of sample sites based on elevation or slope did not vary much, whereas in polygons located in mountainous areas, the sample sites defined based on slope were mostly located in places relatively stable in topography and more representative of the entire polygon. Compared with the environmental covariates histogram peak method used in most of the researches, this method is higher in odds ratio and standard deviation, and the samples defined with this method are bigger in volume of information.【Conclusion】The training samples defined with the geographical neighborhood analysis method based on slope are more representative of the entire polygon than those based on elevation, and contain more information than those defined with the environmental covariates histogram peak method in relevant researches.