Digital soil mapping allows predicting patterns of soil classes on the basis of well-known reference areas and of data mining techniques to model soil-landscape relationships. The purpose of this study was to (1) generate a digital pedological map using data mining techniques to associate geomorphometric and geology variables with soil classes of traditional soil maps in reference areas and (2) validate these maps by different field techniques. The mapping was carried out using the 1:50.000 Botucatu sheet (SF-22-Z-B-VI-3), and 1:50.000 Dois Córregos and São Pedro sheets (São Paulo, Brazil) as reference areas. Training data - soil mapping units (MU) and topographic and geological variables from the reference areas were analyzed by PART, a decision-tree algorithm found on the Weka (Waikato Environment for Knowledge Analysis) software, producing classification rules, which were applied to the Botucatu sheet. Field validation of the produced digital maps was carried out by transect sampling in the zone of São Pedro and by a stratified-random sampling procedure at Botucatu sheet. Accuracy of the mapping unit at São Pedro was 83 %, for the digital and 66 %, for the traditional soil map with simplified legend. Although analysis generated rules for all MU's of the training areas, not all MU's were predicted on the Botucatu sheet, due to differences in relief and geology between training and mapping areas. Digital soil map of Botucatu had an overall accuracy of 52 %, consistent with reconnaissance soil surveys of low intensity, and a kappa index of 0.41, indicating good quality. Larger mapping units on training areas produced more rules, thus reproducing more accurately soil-landscape pattern of the mapped area Validation at the São Pedro sheet by transect sampling suggested that the digital map is cconsistent to high intensity reconnaissance soil surveys; whereas the traditional map (simplified legend) quality corresponded to that of low-intensity soil surveys. Training of the algorithm on maps, not in field-observed points, reduced accuracy of digital soil map by 14 %. Latin hypercube sampling is adequate for mapping areas with large data bases, allows to evaluate the entire mapped area and imparts efficiency to field work. Transect sampling is adequate to evaluate purity of individual mapping units, requires no detailed data base of predictor variables, and allows studies on soil-landscape relationships in pedosequences.
soil-landscape relationship modeling; data mining; quality of soil class maps; latin hypercube sampling; transect sampling; reconnaissance soil surveys
