A Case-based Reasoning Approach to Fuzzy Soil Mapping

has always been a challenge in soil mapping (Hole and Campbell, 1985; Hudson, 1992; McKenzie et al., 2000). Some problems in traditional soil mapping—high cost, high subjecTo date defining the soil–environment relationship for tivity, poor documentation, and low accuracy and precision—have motivated the development of a knowledge-based fuzzy soil mapping soil mapping purpose is still largely a mental process system, named SoLIM (Soil Land Inference Model). The rule-based (Hudson, 1992; McKenzie et al., 2000). A well-trained method of the current SoLIM has its limitations. It requires explicit and experienced soil scientist is capable of properly knowledge of the details of soil–environment relationships and it grasping the soil–environment relationships in a certain assumes that the environmental variables are independent from each area and using these relationships to infer the spatial other. This paper presents a case-based reasoning (CBR) approach distribution of soils over the area. as an alternative to the rule-based method. Case-based reasoning uses Associated with this mental process is the manual knowledge in the form of specific cases to solve a new problem, and process in creating soil maps: with the built soil–landscape the solution is based on the similarities between the new problem model for a mapping area in mind, a soil scientist manuand the available cases. With the CBR method, soil scientists express ally delineates soil polygons on orthophotos under stetheir knowledge by providing locations (cases) indicating the association between a soil and a landscape or environmental configuration. reoscopes. Several problems are associated with this In this way, the soil scientists avoid the difficulties associated with manual process. The first is the high cost (on money, depicting the details of a soil–environment relationship and assuming labor, and time). Zhu et al. (2001) indicated that with the independence of environmental variables. The CBR inference the current rate of soil survey updating, updating all of engine computes the similarity between the environmental configurathe soil surveys in the USA requires 220 yr. The second tion at a given location and that associated with each case representing problem is the high subjectivity. Researchers have noa soil type, and then uses these similarity values to approximate the ticed that different soil scientists may map the same similarity of the local soil at the given location to the given soil type. area in significantly different ways (Bie and Beckett, A case study in southwestern Wisconsin demonstrates that CBR can 1973; Burrough et al., 1997; McBratney and Odeh, 1997; be an easy and effective way for soil scientists to express their knowlMacMillan et al., 2000), and this is at least partially due edge. For the study area, the result from the CBR inference engine is more accurate than that from the traditional soil mapping process. to the inconsistency in the manual mapping process. Case-based reasoning can be a good solution for a knowledge-based Another problem is that the knowledge is hard to prefuzzy soil mapping system. serve in this field and training a qualified soil scientist is expensive. This is because the manual mapping is largely a personal operation that lacks a scheme to guarS mapping is basically an inference process based antee good documentation of the knowledge. Still anon Jenny’s model (Jenny, 1941, 1980). In routine soil other problem is with the polygon-based model. This survey and mapping, this model can be represented as model assumes that the soils are the same everywhere within a polygon and are to be of the type assigned to S f(E) [1] this polygon, and they change abruptly at the polygon where S denotes soil, E denotes environmental variboundary. Apparently, in most situations this assumpables, and f denotes the soil–environment relationship tion is not valid, as soils often change continuously over (soil–landscape model). According to this model, if the both geographical and property spaces (e.g., Burrough environmental conditions at a given location and the et al., 1997; McBratney and Odeh, 1997; Zhu, 1997a). soil–environment relationship are known, then it is posThe manual mapping does not allow this continuous sible to infer the conditions of soil at that given location. variability of soils to be precisely represented, even if With today’s spatial information technologies, including the soil scientists do know the continuous nature of geographic information systems (GIS), remote sensing, soil variation. and the Global Positioning System (GPS), it is possible These problems have motivated the development of to characterize the environmental conditions in details. knowledge-based systems and the application of fuzzy Defining the soil–environment relationship, however, logic in this field. Knowledge-based systems aim at making a good utilization of domain experts’ knowledge, X. Shi, Dep. of Geography, Dartmouth College, 6017 Fairchild, Hanmeanwhile trying to avoid the problems associated with over, NH 03755; A-X. Zhu, State Key Lab of Resources and Environa manual process, such as inconsistency, tediousness, mental Information Systems, Inst. of Geographical Sciences and Natuand loss of knowledge due to personnel change. Reral Resources Res., Chinese Academy of Sciences, Building 917, searchers have used knowledge-based systems to clasDatun Road, An Wai, Beijing 100101, China; A.-X. Zhu, J.E. Burt, and F. Qi, Dep. of Geography, University of Wisconsin-Madison, 550 sify soil samples (Galbraith and Bryant, 1998; Galbraith North Park Street, Madison, WI 53706; D. Simonson, NRCS-USDA, et al., 1998; Holt and Benwell, 1999), predict soil proper1850 Bohmann Drive, Suite C, Richland Center, WI 53581. Received ties (Cook et al., 1996), and map soil-landscape units 25 Apr. 2002. *Corresponding author ([email protected]). Abbreviations: 3D, three-dimensional; CBR, case-based reasoning; Published in Soil Sci. Soc. Am. J. 68:885–894 (2004).  Soil Science Society of America GIS, geographic information systems; SoLIM, soil land inference model. 677 S. Segoe Rd., Madison, WI 53711 USA