Digital elevation models (DEM) are fundamental sources for correlating the occurrence and distribution of soils in the landscape by digital soil mapping (DSM). The influence of the types and resolutions of DEMs on the predictive capability of prediction models of soil classes has not been greatly studied. This study evaluated and compared the effects of different DEMs on the prediction of occurrence of soil mapping units (MU). Twelve predictor variables derived from different DEMs were correlated with the occurrence of MU. The DEMs used were those from the SRTM v4.1, ASTER GDEM v2, TOPODATA, and Brasil em Relevo projects, and the DEMs generated from contour lines on the scale of 1:50,000, with resolutions of 30 and 90 m. The predictive models were guided by a decision tree (Simple Cart) with data sampled at 4,280 random points containing information extracted from soils of a conventional soil map at a scale of 1:20,000 and 12 terrain attributes derived from six DEMs with pixel sizes of 30 and 90 m. The validation of MU prediction models was performed with all the data from the area. The terrain attributes that best explained the occurrence of MU were elevation, slope, flow length, and aspect. The DEMs with 30 m pixel size generated less accurate soil-landscape correlations. The most accurate prediction models with a higher number of estimated MU were generated from the DEMs with spatial resolution of 90 m (SRTM v4.1 and CN90). These are the DEMs most suitable for DSM when flat and slightly rolling terrain predominate.
terrain properties; spatial resolution; pixel size; decision tree
