The information about spatial distribution of soil attributes over a large area is the basic one required for land surface process modeling. Spatial interpolation based on sampling sites in the field is an important way to acquire such information. The existing spatial interpolation methods usually call for sound representativeness of all the sampling sites for the whole study area in soil attribute spatial distribution. However, limited by budget and/or field accessibility for sampling, in most cases, the sampling sites chosen can hardly reflect comprehensively the spatial distribution of soil attributes of the study area. A soil attribute soil map worked out with the existing interpolation methods based on such sampling is often low in precision and the poor representativeness of the sampling makes it hard to measure the uncertainty it brings about in prediction. In order to make reasonable use of such samples and measure prediction uncertainty effectively, this paper presents a new method that uses the individual representativeness of each sample to predict soil attribute of unvisited locations and to quantify prediction uncertainty. Based on the assumption that the more similar the environment conditions of two sites, the more similar the soil attributes of the two, it is held that every sampling site can be used to represent an area similar to the site in environment, and the representativeness can be measured by similarity degree of the two. By analyzing environmental similarity, uncertainty in prediction can be measured, and with environmental similarity as weight, soil attribute values of the area a sampling site represents can be worked out. This method was once used to predict soil organic matter (SOM) content in the surface soil layer of the Yili District, Xinjiang Uyghur Autonomous Region for validation. As demonstrated, the method is an effective approach to using a sampling site that is poor in global representativeness to predict soil attribute spatial distribution of the area it represents, and the uncertainty of the prediction is positively related to the predicated residual, which can be used effectively to indicate credibility of the prediction.