Analysis of Patterns of Pedotransfer Function Estimates: An Approach Based on Classification Trees

Patterns in the errors of three pedotransfer functions (PTFs) for estimating water retention at −10 and −1500 kPa were investigated using a data set of Brazilian soils. Of the three PTFs, one was developed using soils from temperate regions, one using soils from different tropical areas, and one was specifically based on Brazilian soils. We used classification tree models to obtain rules that represent patterns in PTF errors, reflecting differences between measured and predicted water content values. These errors were classified as acceptable or unacceptable according to a defined threshold. The tree models had an accuracy of about 70% and could be used to identify the main attributes that control PTF performance. For the temperate PTF, the main error controls were silt content at −10 kPa and clay content at −1500 kPa. Overall, acceptable errors at one pressure head occurred under the same conditions as the unacceptable errors at the other pressure head. The tree rules reflected the differences between the data sets used to generate the PTFs and the data set used to evaluate the PTFs. Geographical closeness appeared to be insufficient to ensure good performance of a PTF. Rather, the statistical distribution of the input attributes used to develop PTFs plays a more important role in error classification. Classification trees allow us to better understand the factors that control PTF performance and how these factors interact.