【Objective】 Quality of mapping based on prediction of geographic variables is greatly affected by the layout of the sampling sites. Due to the limitation of sampling cost, it is generally expected to have fewer sampling sites that will be able to provide more information for accurate prediction. To achieve such a target of efficient sampling, it is advisable to proceed from the following two point: first, set up sampling sites that are highly representative of the area of interest for better prediction accuracy, and second, reduce the number of sampling sites as many as possible without risking any loss of required accuracy. Based on the assumption that the more similar the two sites in geographic environment, the more similar their geographic elements, it is held that every sampling site can be used to represent an area similar to the site in environment, and the similarity between the sampling site and the sites to be predicted can be used to determine reliability of the prediction, meanwhile, the similarity within the sampling site set can be used to determine redundancy of the sampling site set. So, the layout of efficient sampling sites needs to keep balance between reliability of the prediction and redundancy of the sampling site set. 【Method】 In this paper, a sample differentiation and fusion strategy is set forth for designing of sampling. The differentiation strategy is to increase the number of sampling sites so as to improve reliability of later on predictions, while the fusion strategy is to merge over-similar sampling sites, so as to reduce redundancy of the sampling site set. Through repeated differentiation and fusion, a preset requirement is finally met for prediction reliability and sampling site redundancy. The method has been tested in a case study of a small watershed in Raffelson, Wisconsin of USA. First, a comparative analysis was done of sampling sites varying in prediction reliability with 99 validation sampling sites to determine relationship between prediction reliability and validation accuracy. Then, verification was performed of effectiveness of the proposed strategy improving prediction reliability in its first phase and reducing redundancy of the sampling site set in its second phase. And in the end, comparison was done of the proposed method in this paper with other sampling methods (grid sampling and stratified simple random sampling) using the same number of sampling sites (15, 20 and 25, separately). 【Result】 Results show that prediction reliability is positively related to prediction accuracy, so the former can be used as a better indicator of the latter. From the specific processes of the strategy, it can be discerned that, the differentiation can effectively raise the prediction reliability, while the fusion reduce the redundancy of the sampling site set, and what is more, have little impact on the prediction reliability. The comparisons show that the proposed method is higher in prediction reliability and lower in redundancy, and is 17.3%(n=15), 14.8%(n=20) and 16.2%(n=25) lower than the grid sampling method, and 16.5%(n=15), 15.3%(n=20) and 17.0%(n=25) higher than the stratified simple random sampling method in lowest prediction reliability, respectively, while 8.8%(n=15), 12.8%(n=20) and 20.3%(n=25) lower than the grid sampling method, and 6.4%(n=15), 12.4%(n=20) and 19.6%(n=25) lower than the stratified simple random sampling method, respectively, in redundancy. 【Conclusion】 Therefore, it can be concluded that the proposed method provides a means for obtaining a high prediction reliability and low sampling redundancy in sampling, and hence is a more efficient method for designing sampling schemes than the grid sampling and stratified simple random sampling methods.