The predictive models used in digital soil mapping (DSM) need to be trained with data that most fully capture the variation of terrain and soil properties in order to generate adequate correlations between environmental variables and the occurrence of soil unities. Several methods have been used in DSM to evaluate the accuracy of these models. The aims of this study were to compare the use of three sampling procedures for training a classification and regression tree (CART) algorithm, and evaluate the predictive capacity of the models generated using four methods. The sampling procedures used were: simple random; proportional to the area of each soil mapping unit (MU), and stratified by the number of MUs. The evaluation methods tested were: apparent, percentage division, cross-validation with 10 subsets, and resampling with seven independent data sets. The accuracies of the models estimated by the methods were compared with the measured accuracies. This was achieved by comparing the maps generated, based on each sampling procedure, with the conventional soil map at the scale of 1:50,000. The sampling procedures influenced the number of MUs predicted and the accuracy of the models and of the maps generated. The proportional and stratified sampling procedures resulted in less accurate digital soil maps, and the accuracies of the models varied according to the evaluation method adopted. Random sampling resulted in the most accurate digital soil map and presented accuracy values that were similar for all the evaluation methods tested.
proportional sampling; random sampling; stratified sampling; accuracy evaluation; classification tree
