The increased sophistication of spatial analytical techniques and the availability of digital data about the physical landscape have made it possible to develop techniques for digitally and predictively mapping soil spatial variation to improve not only the speed but also the quality and level of detail of soil survey. Among the many techniques currently under development or under testing, SoLIM (Soil Land Inference Model), a result of a joint research effort by Department of Geography, University of Wisconsin-Madison and Natural Resources Conservation Service, United States Department of Agriculture, is a representative example. This paper examines the current successes, challenges, and future perspectives of digital soil mapping from the experience acquired in developing the SoLIM technology. We hope that this examination will shed some light on critical aspects of digital soil mapping and thereby assist in the ongoing evolution and development of similar technologies. SoLIM was developed to overcome the limitations of current soil survey methods and to increase the efficiency and accuracy of soil resource surveys. The SoLIM approach uses a raster data model coupled with a fuzzy logic-based representation scheme to represent the detail spatial variation of soils. It maps the detailed spatial variation by combining the knowledge, extracted from local soil experts or other sources using artificial intelligence and other machine learning techniques, with data about the soil environment, characterized using GIS and remote sensing techniques. Cases studies in Wisconsin and other test areas in U.S. have shown that the SoLIM approach is much faster than the current approach and the products are about 20-30% more accurate than those produced using the existing methods. In addition, SoLIM can generate a range of products not available using the traditional approach and its products can be easily and continuously updated. There are three major challenges faced by digital soil mapping techniques like SoLIM. The first challenge is how to use the detail information produced by techniques like this. The basic output of SoLIM is a set of fuzzy membership maps of soil classes. The question is how users will use this information in their decision-making process (such as soil interpretation and watershed based modeling). The deployment of these techniques in large part will depend on how readily the products can be used. The second challenge, although much progress has been made in this area, is that how we acquire soil-landscape relationship models and insure the quality of these models, which are needed for predictive mapping, particularly for areas with limited physical accessibility. Third, it has been reported in many studies that predictive soil mapping does not work well in areas with very little relief, mainly because the difference in soils over these area cannot be related to terrain attributes which are the major drivers in predicting mapping. The challenge is then how to develop metrics which will reflect the difference in soils over these gentle/flat areas. There are possible solutions to these challenges. For example, purposive sampling aided by GIS might help us to efficiently develop soil-landscape models over remote areas, fuzzy interpretation might provide a better way for soil interpretation. Data fusion or assimilation techniques may allow us to develop new soil-landscape metrics. However, we believe that the future of these techniques is not solely tied to a set of computer techniques. Rather, the knowledge and the participation of local soil scientists will be critical to the success of these techniques.