Comparing Different Approaches - Data Mining, Geostatistic, and Deterministic Pedology - to Assess the Frequency of WRB Reference Soil Groups in the Italian Soil Regions

The assessment of class frequency in soil map legends is affected by uncertainty, especially at small scales, where generalization is larger. The aim of this study was to test the hypothesis that data mining or geostatistic techniques provide better estimation of class frequency than traditional deterministic pedology in a national soil map. In the map of Italian soil regions compiled at 1:5,000,000 reference scale, soil classes were the WRB Reference Soil Groups (RSGs). Different data mining techniques, namely neural networks, random forests, boosted tree, classification and regression tree, supported vector machine (SVM), were tested and the last one gave the best RSGs predictions, using selected auxiliary variables and 22,015 classified soil profiles. Given the categorical target variable, the multi-collocated indicator cokriging was the algorithm chosen for the geostatistic approach. The first five more frequent RSGs resulting from the three methods were compared. The outcomes were validated with a Bayesian approach on a subset of 10% of geographically representative profiles, kept out before the elaborations. The most frequent classes were uniformly predicted by the three methods, which instead differentiated notably for the classes with a lower occurrence. The Bayesian validation indicated that the SVM method was as reliable as the multi-collocated indicator cokriging, and both more than the deterministic pedological approach. An advantage of the SVM was the possibility to use numeric and categorical variable in the same elaboration, without any previous transformation, which notably reduced the processing time.