Quantitative mapping of soil types based on regression kriging of taxonomic distances with landform and land cover attributes

Nowadays, French soil scientists tend to gather new and existing soil data into a common database. The use of this database potentially allows for resolving environmental issues, largely through soil mapping. The purpose of this study is to present a methodology for mapping soil types illustrated by typical observations in the soil database, in this case from the La Rochelle area on the French Mid-Atlantic Coast. The main hypothesis underlying the method is that soil types result from environmental factors such as landform, parent material, and land cover. The method can be divided into four stages. The first step is to construct a local soil type classification from the database by a two-stage continuous classification procedure. The result of this procedure is that at each observation point, the soil is described by a vector of taxonomic distances to each of k centroidal soil types. In the example given, k=18. The second step involves fitting soil–environment equations, one for each centroidal soil type, by regressing taxonomic distances on layers of multivariate environmental data observed on a fine 20-m grid, by multiple linear regression. In this case, the layers are terrain attributes derived from a digital elevation model and land cover attributes derived from three bands of a SPOT image. The third step is to predict k maps or raster GIS layers representing taxonomic distances to soil types on the 20-m grid, using the soil–environment equations and the kriging of the residuals from the regressions. This results in many potential maps: a summary map depicting the nearest centroidal soil type (the soil type for which the taxonomic distance is least) at each location is possibly the most useful, and another one representing the minimum taxonomic distance which, if considered too large, might suggest locations for further field survey to refine the soil types. A map of standard errors of the kriged taxonomic distance residuals to the nearest centroidal soil type can be made to indicate spatial uncertainty. Continuous fuzzy membership maps can also be constructed from the distances. The fourth step involves validation with an independent soil data set allowing discovery of the nature of the actual prediction errors. Thirty-eight percent of sites in a validation sample of 1234 sites was unequivocally validated, 23% was equivocally validated, and the remainder was predicted wrongly by the method.