In order to study sampling techniques useful for digital soil mapping (DSM), we evaluated the effect of changes in sampling density, based on data from areas already mapped by traditional methods, in regard to the accuracy of decision trees models for generating soil maps using DSM. In two watersheds in northwestern Rio Grande do Sul, Brazil, 1:50,000 scale conventional soils maps were used as reference maps. From the ASTER - GDEM Global Digital Elevation Model and the hydrographic network, maps of predictive variables were generated: elevation, slope, curvature, flow length, flow accumulation, topographic wetness index, and Euclidian distance of the streams. We used random sampling, and tested sampling densities that ranged from 0.1 to 4 points per hectare. Models were trained using Weka software, generating predictive models using different sizes of the final node to obtain decision trees of different sizes. The results indicate that when the size of the decision tree was not controlled, an increase in sampling density resulted in greater overall accuracy in accordance with the basic reference maps and an increase in the number of predicted soil mapping units. When the size of decision trees was controlled, an increase in sampling density did not affect the overall accuracy and had a very slight influence on the number of predicted mapping units.
soil map; GIS; model; prediction
