In the United States, most cooperative soil survey activities are directed toward revision of existing surveys. Such surveys, which are typically decades old, can be thought of as spatial expressions of the surveyor's conception of soil-landscape relations. That is, based on field investigation, laboratory analysis, review of other surveys, etc., the scientist develops soil concepts (soil classes) and associates those concepts with geology, landscape position, slope, curvature, vegetation, and other environmental indicators that can be exploited in mapping. Obviously, those indicators which are directly tied to pedogenesis play a vital role in the soil-landscape model. Traditional survey practice was largely manual, with heavy reliance on stereo aerial photography for line placement (application of the model). There are two primary motivations for survey updates. First, limitations inherent in the manual process prevent totally consistent application of the model, regardless of how complete and well-conceived that model might be. Second, over time soils knowledge improves, and it is obviously desirable to incorporate that knowledge in a revised survey. The existing surveys, though not perfect, contain a wealth of information that can potentially serve as a starting point for a revised survey. Although the scientist's soil-landscape model might not be explicitly documented anywhere, it is implicit in the survey. Our project is motivated by belief that if the model can be recovered from the existing survey, it can be used to jumpstart model development for the new survey. To that end, we have developed a set of data mining tools for extracting soil-landscape relations from a published survey. Map polygons are overlaid on raster data of various sorts, including a number of variables computed from a Digital Elevation Matrix (DEM). The DEM-derived data include slope, aspect, planform and profile curvatures, wetness index, and other terrain indices. Together with geology and other non-DEM data, they define a suite of environmental variables identified by a soil scientist as likely to be important in the study area. Our tools extract knowledge in the form of frequency distributions of pixels within map polygons. That is, for any polygon there is distribution of elevation, slope, etc. indicating the range of environmental conditions over which that polygon has been mapped. By comparing distributions of one map unit with another, we obtain information about how the original surveyor chose to map those units; that is, we find similarities and differences in the environments occupied by the units. By comparing frequency distributions for polygons of the same map unit, we obtain information about the consistency of mapping, and can identify polygons that occupy anomalous environmental settings. This paper describes the data mining tools and their application to soil update for Iowa County in southwestern Wisconsin, U.S.A. We show how the mined data can be used to uncover soil-landscape models, and its utility in both revising soil concepts and suggesting the need for new concepts. We also show that after editing, the mined data can be used in SoLIM--a predictive modeling method--to produce an improved soil survey. We argue that by exploiting recent advances in expert knowledge systems and spatial data processing, methodologies similar to those discussed here have potential for producing faster and more accurate soil survey updates.