Digital Elevation Models (DEM) and satellite images are frequently used as environmental covariates for the prediction of soil classes in Digital Soil Mapping (DSM) procedures. In order to evaluate the input remote sensing image as covariates for soil class prediction, we applied a classification tree (CT) algorithm to predict soil classes at the great group level in a tropical dry forest area with 102 km2 in Brazil. We built 17 CT models using as three sources of DEM (SRTM90m, SRTM30m and Ikonos), two sources of satellite images (Landsat 8 and RapidEye) from two seasons, and airbourne gamma radiometrics images, with different spatial (10-m, 30-m and 90-m) resolutions, as predictors, and, 296 and 128 model training and validation observations, respectively. The results showed that freely available environmental covariates with coarser spatial resolution can produce as good or better great group predictions than more expensive covariates with finer resolution.